Table of Contents

Chapter 7 - Audio Processing

Section 7.1 - Concepts

7.1.1 It's All Audio Processing

We've entitled this chapter "Audio Processing" as if this is a separate topic within the realm of sound. But, actually, everything we do to audio is a form of processing.  Every tool, plug-in, software application, and piece of gear is essentially an audio processor of some sort. What we set out to do in this chapter is to focus on particular kinds of audio processing, covering the basic concepts, applications, and underlying mathematics of these. For the sake of organization, we divide the chapter into processing related to frequency adjustments and processing related to amplitude adjustment, but in practice these two areas are interrelated.

7.1.2 Filters

You have seen in previous chapters how sounds are generally composed of multiple frequency components. Sometimes it’s desirable to increase the level of some frequencies or decrease others. To deal with frequencies, or bands of frequencies, selectively, we have to separate them out. This is done by means of filters. The frequency processing tools in the following sections are all implemented with one type of filter or another.

There are a number of ways to categorize filters. If we classify them according to what frequencies they attenuate, then we have these types of band filters:

  • low-pass filter – retains only frequencies below a given threshold
  • high-pass filter – retains only frequencies above a given threshold
  • bandpass filter – retains only frequencies within a given frequency band
  • bandstop filter – eliminates frequencies within a given frequency band
  • comb filter – attenuates frequencies in a manner that, when graphed in the frequency domain, has a “comb” shape. That is, multiples of some fundamental frequency are attenuated across the audible spectrum
  • peaking filter – boosts or attenuates frequencies in a band
  • shelving filters
    • low-shelf filter – boosts or attenuates low frequencies
    • high-shelf filter – boosts or attenuates high frequencies
Figure 7.1 Frequency responses of different filter types

Figure 7.1 Frequency responses of different filter types

Filters that have a known mathematical basis for their frequency response graphs and whose behavior is therefore predictable at a finer level of detail are sometimes called scientific filters. This is the term Adobe Audition uses for Bessel, Butterworth, Chebyshev, and elliptical filters. The Bessel filter’s frequency response graph is shown in Figure 7.2

Figure 7.2 Bessel scientific filter from Adobe Audition

Figure 7.2 Bessel scientific filter from Adobe Audition

If we classify filters according to the way in which they are designed and implemented, then we have these types:

  • IIR filters – infinite impulse response filters
  • FIR filters – finite impulse response filters

Adobe Audition uses FIR filters for its graphic equalizer but IIR filters for its parametric equalizers (described below.) This is because FIR filters give more consistent phase response, while IIR filters give better control over the cutoff points between attenuated and non-attenuated frequencies. The mathematical and algorithmic differences of FIR and IIR filters are discussed in Section 3. The difference between designing and implementing filters in the time domain vs. the frequency domain is also explained in Section 3.

Convolution filters are a type of FIR filter that can apply reverberation effects so as to mimic an acoustical space. The way this is done is to record a short loud burst of sound in the chosen acoustical space and use the resulting sound samples as a filter on the sound to which you want to apply reverb. This is described in more detail in Section 7.1.6.

7.1.3 Equalization

Audio equalization, more commonly referred to as EQ, is the process of altering the frequency response of an audio signal. The purpose of equalization is to increase or decrease the amplitude of chosen frequency components in the signal. This is achieved by applying an audio filter.

EQ can be applied in a variety of situations and for a variety of reasons. Sometimes, the frequencies of the original audio signal may have been affected by the physical response of the microphones or loudspeakers, and the audio engineer wishes to adjust for these factors. Other times, the listener or audio engineer might want to boost the low end for a certain effect, "even out" the frequencies of the instruments, or adjust frequencies of a particular instrument to change its timbre, to name just a few of the many possible reasons for applying EQ.

Equalization can be achieved by either hardware or software. Two commonly-used types of equalization tools are graphic and parametric EQs. Within these EQ devices, low-pass, high-pass, bandpass, bandstop, low shelf, high shelf, and peak-notch filters can be applied.

7.1.4 Graphic EQ

A graphic equalizer is one of the most basic types of EQ. It consists of a number of fixed, individual frequency bands spread out across the audible spectrum, with the ability to adjust the amplitudes of these bands up or down. To match our non-linear perception of sound, the center frequencies of the bands are spaced logarithmically. A graphic EQ is shown in Figure 7.3. This equalizer has 31 frequency bands, with center frequencies at 20 Hz, 25, Hz, 31 Hz, 40 Hz, 50 Hz, 63 Hz, 80 Hz, and so forth in a logarithmic progression up to 20 kHz. Each of these bands can be raised or lowered in amplitude individually to achieve an overall EQ shape.

While graphic equalizers are fairly simple to understand, they are not very efficient to use since they often require that you manipulate several controls to accomplish a single EQ effect. In an analog graphic EQ, each slider represents a separate filter circuit that also introduces noise and manipulates phase independently of the other filters. These problems have given graphic equalizers a reputation for being noisy and rather messy in their phase response. The interface for a graphic EQ can also be misleading because it gives the impression that you're being more precise in your frequency processing than you actually are. That single slider for 1000 Hz can affect anywhere from one third of an octave to a full octave of frequencies around the center frequency itself, and consequently each actual filter overlaps neighboring ones in the range of frequencies it affects. In short, graphic EQs are generally not preferred by experienced professionals.

Figure 7.3 Graphic EQ in Audacity

Figure 7.3 Graphic EQ in Audacity

7.1.5 Parametric EQ

A parametric equalizer, as the name implies, has more parameters than the graphic equalizer, making it more flexible and useful for professional audio engineering. Figure 7.4 shows a parametric equalizer. The different icons on the filter column show the types of filters that can be applied. They are, from top to bottom, peak-notch (also called bell), low-pass, high-pass, low shelf, and high shelf filters. The available parameters vary according to the filter type. This particular filter is applying a low-pass filter on the fourth band and a high-pass filter on the fifth band.

Figure 7.4 Parametric EQ in Cakewalk Sonar

Figure 7.4 Parametric EQ in Cakewalk Sonar

Aside:  The term "paragraphic EQ" is used for a combination of a graphic and parametric EQ, with sliders to change amplitudes and parameters that can be set for Q, cutoff frequency, etc.

For the peak-notch filter, the frequency parameter corresponds to the center frequency of the band to which the filter is applied. For the low-pass, high-pass, low-shelf, and high-shelf filters, which don’t have an actual “center,” the frequency parameter represents the cut-off frequency. The numbered circles on the frequency response curve correspond to the filter bands. Figure 7.5 shows a low-pass filter in band 1 where the 6 dB down point – the point at which the frequencies are attenuated by 6 dB – is set to 500 Hz.

Figure 7.5 Low-pass filter in a parametric EQ with cut-off frequency of 500 Hz

Figure 7.5 Low-pass filter in a parametric EQ with cut-off frequency of 500 Hz

The gain parameter is the amount by which the corresponding frequency band is boosted or attenuated. The gain cannot be set for low or high-pass filters, as these types of filters are designed to eliminate all frequencies beyond or up to the cut-off frequency.

The Q parameter is a measure of the height vs. the width of the frequency response curve. A higher Q value creates a steeper peak in the frequency response curve compared to a lower one, as shown in Figure 7.6.

Some parametric equalizers use a bandwidth parameter instead of Q to control the range of frequencies for a filter. Bandwidth works inversely from Q in that a larger bandwidth represents a larger range of frequencies. The unit of measurement for bandwidth is typically an octave. A bandwidth value of 1 represents a full octave of frequencies between the 6 dB down points of the filter.

Figure 7.6 Comparison of Q values for two peak filters

Figure 7.6 Comparison of Q values for two peak filters

7.1.6 Reverb


When you work with sound either live or recorded, the sound is generally captured with the microphone very close to the source of the sound. With the microphone very close, and particularly in an acoustically treated studio with very little reflected sound, it is often desired or even necessary to artificially add a reverberation effect to create a more natural sound, or perhaps to give the sound a special effect. Usually a very dry initial recording is preferred, so that artificial reverberation can be applied more uniformly and with greater control.

There are several methods for adding reverberation. Before the days of digital processing this was accomplished using a reverberation chamber. A reverberation chamber is simply a highly reflective, isolated room with very low background noise. A loudspeaker is placed at one end of the room and a microphone is placed at the other end. The sound is played into the loudspeaker and captured back through the microphone with all the natural reverberation added by the room. This signal is then mixed back into the source signal, making it sound more reverberant. Reverberation chambers vary in size and construction, some larger than others, but even the smallest ones would be too large for a home, much less a portable studio.

Because of the impracticality of reverberation chambers, most artificial reverberation is added to audio signals using digital hardware processors or software plug-ins, commonly called reverb processors. Software digital reverb processors use software algorithms to add an effect that sounds like natural reverberation. These are essentially delay algorithms that create copies of the audio signal that get spread out over time and with varying amplitudes and frequency responses.

A sound that is fed into a reverb processor comes out of that processor with thousands of copies or virtual reflections. As described in Chapter 4, there are three components of a natural reverberant field. A digital reverberation algorithm attempts to mimic these three components.

The first component of the reverberant field is the direct sound. This is the sound that arrives at the listener directly from the sound source without reflecting from any surface. In audio terms, this is known as the dry or unprocessed sound. The dry sound is simply the original, unprocessed signal passed through the reverb processor. The opposite of the dry sound is the wet or processed sound. Most reverb processors include a wet/dry mix that allows you to balance the direct and reverberant sound. Removing all of the dry signal leaves you with a very ambient effect, as if the actual sound source was not in the room at all.

The second component of the reverberant field is the early reflections. Early reflections are sounds that arrive at the listener after reflecting from the first one or two surfaces. The number of early reflections and their spacing vary as a function of the size and shape of the room. The early reflections are the most important factor contributing to the perception of room size. In a larger room, the early reflections take longer to hit a wall and travel to the listener. In a reverberation processor, this parameter is controlled by a pre-delay variable. The longer the pre-delay, the longer time you have between the direct sound and the reflected sound, giving the effect of a larger room. In addition to pre-delay, controls are sometimes available for determining the number of early reflections, their spacing, and their amplitude. The spacing of the early reflections indicates the location of the listener in the room. Early reflections that are spaced tightly together give the effect of a listener who is closer to a side or corner of the room. The amplitude of the early reflections suggests the distance from the wall. On the other hand, low amplitude reflections indicate that the listener is far away from the walls of the room.

The third component of the reverberant field is the reverberant sound. The reverberant sound is made of up all the remaining reflections that have bounced around many surfaces before arriving at the listener. These reflections are so numerous and close together that they are perceived as a continuous sound. Each time the sound reflects off a surface, some of the energy is absorbed. Consequently, the reflected sound is quieter than the sound that arrives at the surface before being reflected. Eventually all the energy is absorbed by the surfaces and the reverberation ceases. Reverberation time is the length of time it takes for the reverberant sound to decay by 60 dB, effectively a level so quiet it ceases to be heard. This is sometimes referred to as the RT60, or also the decay time. A longer decay time indicates a more reflective room.

Because most surfaces absorb high frequencies more efficiently than low frequencies, the frequency response of natural reverberation is typically weighted toward the low frequencies. In reverberation processors, there is usually a parameter for reverberation dampening. This applies a high shelf filter to the reverberant sound that reduces the level of the high frequencies. This dampening variable can suggest to the listener the type of reflective material on the surfaces of the room.

Figure 7.7 shows a popular reverberation plug-in. The three sliders at the bottom right of the window control the balance between the direct, early reflection, and reverberant sound. The other controls adjust the setting for each of these three components of the reverberant field.

Figure 7.7 The TrueVerb reverberation plug-in from Waves

Figure 7.7 The TrueVerb reverberation plug-in from Waves

The reverb processor pictured in Figure 7.8 is based on a complex computation of delays and filters that achieve the effects requested by its control settings. Reverbs such as these are often referred to as algorithmic reverbs, after their unique mathematical designs.

Aside:  Convolution is a mathematical process that operates in the time-domain – which means that the input to the operation consists of the amplitudes of the audio signal as they change over time. Convolution in the time-domain has the same effect as mathematical filtering in the frequency domain, where the input consists of the magnitudes of frequency components over the frequency range of human hearing. Filtering can be done in either the time domain or the frequency domain, as will be explained in Section 3.

There is another type of reverb processor called a convolution reverb, which creates its effect using an entirely different process. A convolution reverb processor uses an impulse response (IR) captured from a real acoustic space, such as the one shown in Figure 7.8. An impulse response is essentially the recorded capture of a sudden burst of sound as it occurs in a particular acoustical space. If you were to listen to the IR, which in its raw form is simply an audio file, it would sound like a short “pop” with somewhat of a unique timbre and decay tail. The impulse response is applied to an audio signal by a process known as convolution, which is where this reverb effect gets its name. Applying convolution reverb as a filter is like passing the audio signal through a representation of the original room itself. This makes the audio sound as if it were propagating in the same acoustical space as the one in which the impulse response was originally captured, adding its reverberant characteristics.

With convolution reverb processors, you lose the extra control provided by the traditional pre-delay, early reflections, and RT60 parameters, but you often gain a much more natural reverberant effect. Convolution reverb processors are generally more CPU intensive than their more traditional counterparts, but with the speed of modern CPUs, this is not a big concern. Figure 7.8 shows an example of a convolution reverb plug-in.

Figure 7.8 A convolution reverb processor from Logic

Figure 7.8 A convolution reverb processor from Logic

7.1.7 Flange

Flange is the effect of combing out frequencies in a continuously changing frequency range. The flange effect is created by adding two identical audio signals, with one slightly delayed relative to the other, usually on the order of milliseconds. The effect involves continuous changes in the amount of delay, causing the combed frequencies to sweep back and forth through the audible spectrum.

In the days of analog equipment like tape decks, flange was created mechanically in the following manner: Two identical copies of an audio signal (usually music) were played, simultaneously and initially in sync, on two separate tape decks. A finger was pressed slightly against the edge (called the flange) of one of the tapes, slowing down its rpms. This delay in one of the copies of the identical waveforms being summed resulted in the combing out of a corresponding fundamental frequency and its harmonics. If the pressure increased continuously, the combed frequencies swept continuously through some range. When the finger was removed, the slowed tape would still be playing behind the other. However, pressing a finger against the other tape could sweep backward through the same range of combed frequencies and finally put the two tapes in sync again.

Artificial flange can be created through mathematical manipulation of the digital audio signal. However, to get a classic sounding flanger, you need to do more than simply delay a copy of the audio. This is because tape decks used in analog flanging had inherent variability that caused additional phase shifts and frequency combing, and thus they created a more complex sound. This fact hasn’t stopped clever software developers, however. The flange processor shown in Figure 7.9 from Waves is one that includes a tape emulation mode and includes presets that emulate several kinds of vintage tape decks and other analog equipment.

Figure 7.9 A digital flange processor

Figure 7.9 A digital flange processor

7.1.8 Vocoders

A vocoder (voice encoder) is a device that was originally developed for low bandwidth transmission of voice messages, but is now used for special voice effects in music production.

The original idea behind the vocoder was to encode the essence of the human voice by extracting just the most basic elements – the consonant sounds made by the vocal chords and the vowel sounds made by the modulating effect of the mouth. The consonants serve as the carrier signal and the vowels (also called formants) serve as the modulator signal.  By focusing on the most important elements of speech necessary for understanding, the vocoder encoded speech efficiently, yielding a low bandwidth for transmission. The resulting voice heard at the other end of the transmission didn't have the complex frequency components of a real human voice, but enough information was there for the words to be intelligible.

Today’s vocoders, used in popular music, combine voice and instruments to make the instrument sound as if it’s speaking, or conversely, to make a voice have a robotic or “techno” sound. The concept is still the same, however. Harmonically-rich instrumental music serves as the carrier, and a singer’s voice serves as the modulator. An example of a software vocoder plug-in is shown in Figure 7.10.

Figure 7.10 A vocoder processor

Figure 7.10 A vocoder processor

7.1.9 Autotuners

An autotuner is a software or hardware processor that is able to move a pitch of the human voice to the frequency of the nearest desired semitone. The original idea was that if the singer was slightly off-pitch, the autotuner could correct the pitch. For example, if the singer was supposed to be on the note A at a frequency of 440 Hz, and she was actually singing the note at 435 Hz, the autotuner would detect the discrepancy and make the correction.

Aside:  Autotuners have also been used in popular music as an effect rather than a pitch correction. Snapping a pitch to set semitones can create a robotic or artificial sound that adds a new complexion to a song. Cher used this effect in her 1998 Believe album. In the 2000s, T-Pain further popularized its use in R&B and rap music.

If you think about how an autotuner might be implemented, you'll realize the complexities involved. Suppose you record a singer singing just the note A, which she holds for a few seconds. Even if she does this nearly perfectly, her voice contains not just the note A but harmonic overtones that are positive integer multiples of the fundamental frequency. Your algorithm for the software autotuner first must detect the fundamental frequency – call it f – from among all the harmonics in the singer's voice. It then must determine the actual semitone nearest to f. Finally, it has to move f and all of its harmonics by the appropriate adjustment. All of this sounds possible when a single clear note is steady and sustained long enough for your algorithm to analyze it. But what if your algorithm has to deal with a constantly-changing audio signal, which is the nature of music? Also, consider the dynamic pitch modulation inherent in a singer’s vibrato, a commonly used vocal technique. Detecting individual notes, separating them one from the next, and snapping each sung note and all its harmonics to appropriate semitones is no trivial task. An example of an autotune processor is shown in Figure 7.11.

Figure 7.11 An autotune processor

Figure 7.11 An autotune processor

7.1.10 Dynamics Processing Amplitude Adjustment and Normalization

One of the most straightforward types of audio processing is amplitude adjustment – something as simple as turning up or down a volume control. In the analog world, a change of volume is achieved by changing the voltage of the audio signal. In the digital world, it's achieved by adding to or subtracting from the sample values in the audio stream – just simple arithmetic.

An important form of amplitude processing is normalization, which entails increasing the amplitude of the entire signal by a uniform proportion. Normalizers achieve this by allowing you to specify the maximum level you want for the signal, in percentages or dB, and increasing all of the samples’ amplitudes by an identical proportion such that the loudest existing sample is adjusted up or down to the desired level. This is helpful in maximizing the use of available bits in your audio signal, as well as matching amplitude levels across different sounds. Keep in mind that this will increase the level of everything in your audio signal, including the noise floor.

Figure 7.12 Normalizer from Adobe Audition

Figure 7.12 Normalizer from Adobe Audition Dynamics Compression and Expansion

Dynamics processing refers to any kind of processing that alters the dynamic range of an audio signal, whether by compressing or expanding it. As explained in Chapter 5, the dynamic range is a measurement of the perceived difference between the loudest and quietest parts of an audio signal. In the case of an audio signal digitized in n bits per sample, the maximum possible dynamic range is computed as the logarithm of the ratio between the loudest and the quietest measurable samples – that is, 20\log_{10}\left ( \frac{2^{n-1}}{1/2} \right )dB. We saw in Chapter 5 that we can estimate the dynamic range as 6n dB. For example, the maximum possible dynamic range of a 16-bit audio signal is about 96 dB, while that of an 8-bit audio signal is about 48 dB.

The value of 20\log_{10}\left ( \frac{2^{n-1}}{1/2} \right )dB gives you an upper limit on the dynamic range of a digital audio signal, but a particular signal may not occupy that full range. You might have a signal that doesn't have much difference between the loudest and quietest parts, like a conversation between two people speaking at about the same level. On the other hand, you might have at a recording of a Rachmaninoff symphony with a very wide dynamic range. Or you might be preparing a background sound ambience for a live production. In the final analysis, you may find that you want to alter the dynamic range to better fit the purposes of the recording or live performance. For example, if you want the sound to be less obtrusive, you may want to compress the dynamic range so that there isn't such a jarring effect from a sudden difference between a quiet and a loud part.

In dynamics processing, the two general possibilities are compression and expansion, each of which can be done in the upwards or downwards direction (Figure 7.13). Generally, compression attenuates the higher amplitudes and boosts the lower ones, the result of which is less difference in level between the loud and quiet parts, reducing the dynamic range. Expansion generally boosts the high amplitudes and attenuates the lower ones, resulting in an increase in dynamic range. To be precise:

  • Downward compression attenuates signals that are above a given threshold, not changing signals below the threshold. This reduces the dynamic range.
  • Upward compression boosts signals that are below a given threshold, not changing signals above the threshold. This reduces the dynamic range.
  • Downward expansion attenuates signals that are below a given threshold, not changing signals above the threshold. This increases the dynamic range.
  • Upward expansion boosts signals that are above a given threshold, not changing signals below the threshold. This increases the dynamic range.

The common parameters that can be set in dynamics processing are the threshold, attack time, and release time. The threshold is an amplitude limit on the input signal that triggers compression or expansion. (The same threshold triggers the deactivation of compression or expansion when it is passed in the other direction.) The attack time is the amount of time allotted for the total amplitude increase or reduction to be achieved after compression or expansion is triggered. The release time is the amount of time allotted for the dynamics processing to be "turned off," reaching a level where a boost or attenuation is no longer being applied to the input signal.

Figure 7.13 Dynamics compression and expansion

Figure 7.13 Dynamics compression and expansion

Adobe Audition has a dynamics processor with a large amount of control. Most dynamics processor's controls are simpler than this – allowing only compression, for example, with the threshold setting applying only to downward compression. Audition's processor allows settings for compression and expansion and has a graphical view, and thus it's a good one to illustrate all of the dynamics possibilities.

Figure 7.14 shows two views of Audition's dynamics processor, the graphic and the traditional, with settings for downward and upward compression. The two views give the same information but in a different form.

In the graphic view, the unprocessed input signal is on the horizontal axis, and the processed input signal is on the vertical axis. The traditional view shows that anything above -35 dBFS should be compressed at a 2:1 ratio. This means that the level of the signal above -35 dBFS should be reduced by ½ . Notice that in the graphical view, the slope of the portion of the line above an input value of -35 dBFS is ½. This slope gives the same information as the 2:1 setting in the traditional view. On the other hand, the 3:1 ratio associated with the -55 dBFS threshold indicates that for any input signal below -55 dBFS, the difference between the signal and -55 dBFS should be reduced to 1/3 the original amount. When either threshold is passed (-35 or -55 dBFS), the attack time (given on a separate panel not shown) determines how long the compressor takes to achieve its target attenuation or boost. When the input signal moves back between the values of -35 dBFS and -55 dBFS, the release time determines how long it takes for the processor to stop applying the compression.

Figure 7.14 Dynamics processing in Adobe Audition, downward and upward compression

Figure 7.14 Dynamics processing in Adobe Audition, downward and upward compression

A simpler compressor – one of the ARDOUR LADSPA plug-ins, is shown in Figure 7.15. In addition to attack, release, threshold, and ratio controls, this compressor has knee radius and makeup gain settings. The knee radius allows you to shape the attack of the compression to something other than linear, giving a potentially smoother transition when it kicks in. The makeup gain setting (often called simply gain) allows you to boost the entire output signal after all other processing has been applied.

Figure 7.15 SC1 Compressor plug-in for Ardour

Figure 7.15 SC1 Compressor plug-in for Ardour Limiting and Gating

Aside:  A limiter could be thought of as a compressor with a compression ratio of infinity to 1.  See the next section on dynamics compression.

A limiter is a tool that prevents the amplitude of a signal from going over a given level. Limiters are often applied on the master bus, usually post-fader. Figure 7.16 shows the LADSPA Fast Lookahead Limiter plug-in. The input gain control allows you to increase the input signal before it is checked by the limiter. This limiter looks ahead in the input signal to determine if it is about to go above the limit, in which case the signal is attenuated by the amount necessary to bring it back within the limit. The lookahead allows the attenuation to happen almost instantly, and thus there is no attack time. The release time indicates how long it takes to go back to 0 attenuation when limiting the current signal amplitude is no longer necessary. You can watch this work in real-time by looking at the attenuation slider on the right, which bounces up and down as the limiting is put into effect.

Figure 7.16 Limiter LADSPA plug-in

Figure 7.16 Limiter LADSPA plug-in

A gate allows an input signal to pass through only if it is above a certain threshold. A hard gate has only a threshold setting, typically a level in dB above or below which the effect is engaged. Other gates allow you to set an attack, hold, and release time to affect the opening, holding, and closing of the gate (Figure 7.18). Gates are sometimes used for drums or other instruments to make their attacks appear sharper and reduce the bleed from other instruments unintentionally captured in that audio signal.

Figure 7.18 Gate (Logic Pro)

Figure 7.18 Gate (Logic Pro)

A noise gate is a specially designed gate that is intended to reduce the extraneous noise in a signal. If the noise floor is estimated to be, say, -80 dBFS, then a threshold can be set such that anything quieter than this level is blocked out, effectively transmitted as silence. A hysteresis control on a noise gate indicates that there is a threshold difference between opening and closing the gate. In the noise gate in Figure 7.18, the threshold of -50 dB and the hysteresis setting of -3 dB indicate that the gate closes at -50 dBFS and opens again at -47 dBFS. The side chain controls allow some signal other than the main input signal to determine when the input signal is gated. The side chain signal could cause the gate to close based on the amplitudes of only the high frequencies (high cut) or low frequencies (low cut).

In a practical sense, there is no real difference between a gate and a noise gate. A common misconception is that noise gates can be used to remove noise in a recording. In reality all they can really do is mute or reduce the level of the noise when only the noise is present. Once any part of the signal exceeds the gate threshold, the entire signal is allowed through the gate, including the noise. Still, it can be very effective at clearing up the audio in between words or phrases on a vocal track, or reducing the overall noise floor when you have multiple tracks with active regions but no real signal, perhaps during an instrumental solo.

Figure 7.18 Noise gate (Logic Pro)

Figure 7.18 Noise gate (Logic Pro)

Section 7.2 - Applications

7.2.1 Mixing Mixing Contexts and Devices

Aside:  The fact that digital consoles often follow analog models of control and layout is somewhat of a hot topic. On one hand, this similarity provides some standardization and ease of transition between the two types of consoles. Yet with all of the innovations in user interface technology, you might wonder why these implementations have remained so “old fashioned.” Many people are beginning to use hi-tech UI devices like the iPad along with wireless control protocols like OSC to reinvent the way mixing and audio manipulation is done. While it may take some time for these new techniques to emerge and catch on, the possibilities they provide are both fascinating and seemingly limitless.

A mixing console, or mixer, is a device that takes several different audio signals and mixes them together, to be sent to another device in a more consolidated or organized manner. Mixing can be done in a variety of contexts. Mixing during a live performance requires that an audio engineer balance the sounds from a number of sources. Mixing is also done in the sound studio, as the recordings from multiple channels or on multiple tracks are combined.

Mixing can also be done with a variety of tools. An audio engineering doing the mixing of a live performance could use a hardware device like the one shown in Figure 7.19, an analog mixing console. Digital mixers have now become more common (Figure 7.20), and as you can see, they look pretty much the same as their analog counterparts. Software mixers, with user interfaces modeled after equivalent hardware, are a standard part of audio processing programs like Pro Tools, Apple Logic, Ableton Live, and Cakewalk Sonar. The mixing view for a software mixer is sometimes called the console view, as is the case with Cakewalk Sonar, pictured in Figure 7.21.

Figure 7.19 Analog mixing console

Figure 7.19 Soundcraft K2 Analog mixing console

Figure 7.20 A digital mixing console

Figure 7.20 A Yamaha LS9 digital mixing console

In the following section, we introduce the different components and functions of mixers. Whether a mixer is analog or digital, hardware or software, is not the point. The controls and functions of mixers are generally the same no matter what type you're dealing with or the context in which you're doing the mixing.  A Max demo on mixing consoles is included in Chapter 8.

Figure 7.21 Console view (mixing view) in Cakewalk Sonar

Figure 7.21 Console view (mixing view) in Cakewalk Sonar Inputs and Outputs

The original concept behind a mixer was to take the signals from multiple sources and combine them into a single audio signal that could be sent to a recording device or to an amplification system in a performance space. These so-called “mix down” consoles would have several audio input connections but very few output connections. With the advent of surround sound, distributed sound reinforcement systems, multitrack recorders, and dedicated in-ear monitors, most modern mixing consoles have just as many, if not more, outputs than inputs, allowing the operator to create many different mixes that are delivered to different destinations.

Consider the situation of a recording session of a small rock band. You could easily have more than twenty-four microphones spread out across the drums, guitars, vocalists, etc. Each microphone connects to the mixing console on a separate audio input port and is fed into an input channel on the mixing console. Each channel has a set of controls that allows you to optimize and adjust the volume level and frequency response of the signal and send that signal to several output channels on the mixing console. Each output channel represents a different mix of the signals from the various microphones. The main mix output channel likely contains a mix of all the different microphones and is sent to a pair (or more) of monitor loudspeakers in the control room for the recording engineer and other participants to listen to the performance from the band. This main mix may also represent the artistic arrangement of the various inputs, decided upon by the engineer, producer, and band members, eventually intended for mixed-down distribution as a stereo or surround master audio file. Each performer in the band is also often fed a separate auxiliary output mix into her headphones. Each auxiliary mix contains a custom blend of the various instruments that each musician needs to hear in order to play her part in time and in tune with the rest of the band. Ideally, the actual recording is not a mix at all. Instead, each input channel has a direct output connection that sends the microphone signal into a dedicated channel on a multitrack recording device, which in the digital age is often a dedicated computer DAW. This way the raw, isolated performances are captured in their original state, and the artistic manipulation of the signals can be accomplished incrementally and non-destructively during the mixing process. Channel Strips

Configuring all the knobs, buttons, and faders on a suitably sized mixing console makes all of the above functions possible. When you see a large mixing console like the one pictured in Figure 7.19, you might feel intimidated by all the knobs and buttons. It’s important to realize that most of the controls are simply duplicates. Each input channel is represented by a vertical column, or channel strip, of controls as shown in Figure 7.22.

It’s good to realize that the audio signal typically travels through the channel strip and its various controls from top to bottom. This makes it easy to visualize the audio signal path and understand how and when the audio signal is being affected. For example, you’ll usually find the preamp gain control at the top of the channel strip, as this is the first circuit the audio signal encounters, while the level fader at the bottom is the last component the signal hits as it leaves the channel strip to be mixed with the rest of the individual signals.

Figure 7.22 A single channel strip from an analog mixing console

Figure 7.22 A single channel strip from the Soundcraft K2 analog mixing console Input Connectors

Each input channel has at least one input connector, as shown in Figure 7.23. Typically this is an XLR connector. Some mixing consoles also have a ¼" TRS connector on each input channel. The idea for including both is to use the XLR connector for microphone signals and the ¼" connector for line level or high impedance instrument signals, though you can’t use both at the same time. In some cases, both connectors feed into the same input circuitry, allowing you to use the XLR connector for line level signals as well as microphone signals. This is often desirable, and whenever possible you should use the XLR connector rather than the ¼" because of its benefits such as a locking connection. In some cases, the ¼" connector feeds into the channel strip on a separate path from the XLR connector, bypassing the microphone preamplifier or encountering a -20 dB attenuation before entering the preamplifier. In this situation, running a line level signal through the XLR connector may result in a clipped signal because there is no gain adjustment to compensate for the increased voltage level of the line level signal. Each mixing console implements these connectors differently, so you’ll need to read the manual to find out the specific configuration and input specifications for your mixing console.

Figure 7.23 Input connectors for a single channel on a mixing console

Figure 7.23 Input connectors for a single channel on the Soundcraft K2 mixing console Gain Section

The gain section of the channel strip includes several controls. The most important is the gain knob. Sometimes labeled trim, this knob controls the preamplifier for the input channel. The preamplifier is an electrical circuit that can amplify the incoming audio signal to the optimal line level voltage suitable for use within the rest of the console. The preamplifier is often designed for high quality and very low noise so that it can boost the audio signal without adding a lot of noise or distortion. Because of the sheer number of electrical circuits an audio signal can pass through in a mixing console, the signal can pick up a lot of noise as it travels around in the console. The best way to minimize the effects of this noise is to increase the signal-to-noise ratio from the very start. Since the preamplifier is able to increase the level of the incoming audio signal without increasing the noise level in the console, you can use the preamplifier to increase the ratio between the noise floor of the mixing console and the level of your audio signal. Therefore, the goal of the gain knob is to achieve the highest value possible without clipping the signal. (Chapter 8 has more details on gain setting.)

Figure 7.24 Gain section of an input channel strip

Figure 7.24 Gain section of an input channel strip on the Soundcraft K2 analog mixing console

This is the only place in the console (and likely your entire sound system) where you can increase the level of the signal without also increasing the noise. Thus, you should get all the gain you can at this stage. You can always turn the level down later in the signal chain. Don’t succumb to the temptation to turn down the mixing console preamplifier as a convenient way to fix problems caused downstream by power amplifiers and loudspeakers that are too powerful or too sensitive for your application. Also, you should not turn down the preamplifier in an effort to get all the channel faders to line up in a straight row. These are excellent ways to create a noisy sound system because you're decreasing the signal-to-noise ratio for the incoming audio signal. Once you’ve set that gain knob to the highest level you can without clipping the signal, the only reason you should ever touch it again is if the signal coming in to the console gets louder and starts clipping the input.

If you're feeding a line level signal into the channel, you might find that you're clipping the signal even though the gain knob is turned all the way down. Most mixing consoles have a pad button next to the gain knob. This pad button (sometimes labeled “-20 dB”, “Line”, “range” or “Mic/Line”) attenuates the signal by -20 dB, which should allow you to find a setting on your gain knob that doesn’t clip. Using the pad button shouldn’t necessarily be something you do automatically when using line level signals, as you’re essentially undoing 20 dB of built-in signal-to-noise ratio. Don’t use it unless you have to. Be aware that sometimes this button also serves to reroute the input signal using the ¼" input instead of the XLR. On some consoles that have both ¼" and XLR inputs yet don’t have a pad button, it’s because the -20 dB attenuation is already built in to the signal chain of the ¼" input. These are all factors to consider when deciding how to connect your equipment to the mixing console. (A Max demo in Chapter 8 provides more information on gain setting.)

Another button you'll commonly find next to the gain knob is labeled Ø. This is probably the most misunderstood button in the world of sound. Unfortunately, the mixing console manufacturers contribute to the confusion by labeling this button with the universal symbol for phase. In reality, this button has nothing to do with phase. This is a polarity button. Pressing this button simply inverts the polarity of your signal.

The badly-chosen symbol for the polarity button is inherited from the general confusion among sound practitioners about the difference between phase and polarity. It's true that for pure sine waves, a 180-degree phase shift is essentially identical to a polarity inversion. But that's the only case where these two concepts intersect. In the real world of sound, pure sine waves are hardly ever encountered. For complex sounds that you deal with in practice, phase and polarity are fundamentally different. Phase changes in complex sounds are typically the result of an offset in time. The phase changes as a result of timing offsets are not consistent across the frequency spectrum. A shift in time that would create a 180-degree phase offset for a 1 kHz sound would create a 360-degree phase offset for 2 kHz. This inconsistent phase shift across the frequency spectrum for complex sounds is the cause of comb filtering when two identical sounds are mixed together with an offset in time. Given that a mixing console is all about mixing sounds, it is very easy to cause comb filtering when mixing two microphones that are picking up the same sound at two different distances resulting in a time offset. If you think the button in question adjusts the phase of your signal (as the symbol on the button suggests), you might come to the conclusion that pressing this button manipulates the timing of your signal and compensate for comb filter problems. Nothing could be further from the truth. In a comb filter situation, pressing the polarity button for one of the two signals in question simply convert all cancelled frequencies into frequencies that reinforce each other. All the frequencies that were reinforcing each other will now cancel out. Once you’ve pressed this button, you still have a comb filter. It’s just an inverted comb filter. When you encounter two channels on your console that cause a comb filter when mixed together, a better strategy is to simply eliminate one of the two signals. After all, if these two signals are identical enough to cause a comb filter, you don’t really need both of them in your mix, do you? Simply ducking the fader on one of the two channels will solve your comb filter problem much more efficiently, and certainly more so than using the polarity button.

If this button has nothing to do with phase, what reason could you possibly have to push it? There are many situations where you might run into a polarity problem with one of your input signals. The most common is the dreaded “pin 3 hot” problem. In Chapter 1, we talked about the pinout for an XLR connector. We said that pin 2 carries the positive or “hot” signal and pin 3 carries the negative or “cold” signal. This is a standard from the Audio Engineering Society that was ratified in 1982. Prior to that, each manufacturer did things differently. Some used pin 2 as hot and some used pin 3 as hot. This isn’t really a problem until you start mixing and matching equipment from different manufacturers. Let’s assume your microphone uses pin 2 as hot, but your mixing console uses pin 3 as hot. In that situation, the polarity of the signal coming into the mixing console is inverted. Now if you connect another microphone to a second channel on your mixing console and that microphone also uses pin 3 as hot, you have two signals in your mixing console that are running in opposite polarity. In these situations, having a polarity button on each channel strip is an easy way to solve this problem. Despite the pin 2 hot standard being now thirty years old, there are still some manufacturers making pin-3-hot equipment.

Even if all your equipment is running pin 2 hot, you could still have a polarity inversion happening in your cables. If one end of your cable is accidentally wired up incorrectly (which happens more often than you might think), you could have a polarity inversion when you use that cable. You could take the time to re-solder that connector (which you should ultimately take care of), but if time is short or the cable is hard to get to, you could simply press the polarity button on the mixing console and instantly solve the problem.

There could be artistic reasons you would want to press the polarity button. Consider the situation where you are trying to capture the sound of a drum. If you put the microphone over the top of the drum, when the drum is hit, the diaphragm of the microphone pulls down towards the drum. When this signal passes through your mixing console on to your loudspeakers, the loudspeaker driver also pulls back away from you. Wouldn’t it make more sense for the loudspeaker driver to jump out towards you when the drum is hit? To solve this problem you could go back to the drummer and move the microphone so it sits underneath the drum, or you could save yourself the trip and just press the polarity button. The audible difference here might be subtle, but when you put enough subtle differences together, you can often get a significant difference in audio quality.

Another control commonly found in the gain section is the phantom power button. Phantom power is a 48-volt electrical signal that is sent down the shield of the microphone cable to power condenser microphones. In our example, there is a dedicated 48-volt phantom power button for each input channel strip. In some consoles, there's a global phantom power button that turns on phantom power for all inputs.

The last control that is commonly found in the gain section of the console is a high-pass filter. Pressing this button filters out frequencies below the cutoff frequency for the filter. Sometimes this button has a fixed cutoff frequency of 80Hz, 100Hz, or 125Hz. Some mixing consoles give you a knob along with the button that allows you to set a custom cutoff frequency for the high pass filter. When working with microphones, it's very easy to pick up unwanted sounds that have nothing to do with the sound you’re trying to capture. Footsteps, pops, wind, and handling noise from people touching and moving the microphone are all examples of unwanted sounds that can show up in your microphone. The majority of these sounds fall in very low frequencies. Most musical instruments and voices do not generate frequencies below 125 Hz, so you can safely use a high-pass to filter out frequencies lower than that. Engaging this filter removes most of these unwanted sounds before they enter the signal chain in your system without affecting the good sounds you’re trying to capture. Still, all filters have an effect on the phase of the frequencies surrounding the cutoff frequency, and they can introduce a small amount of additional noise into the signal. For this reason, you should leave the high-pass filter disengaged unless you need it. Insert

Next to the channel input connectors typically there is a set of insert connections. Insert connections consist of an output and input that allow you to connect some kind of external processing device in line with the signal chain in the channel strip. The insert output takes the audio signal from the channel directly after it exits the preamplifier, though some consoles let you choose at what point in the signal path the insert path lies. Thinking back to the top-down signal flow, the insert connections are essentially “inserting” an extra component at that point on the channel strip. In this case, the component isn’t built into the channel strip like the EQ or pan controls. Rather, the device is external and can be whatever the engineer wishes to use. If, for example, you want to compress that dynamics of the audio on input channel 1, you can connect the insert output from channel 1 to the input of an external compressor. Then the output of the compressor can be connected to the insert input on channel 1 of the mixing console. The compressed signal is then fed back into the channel strip and continues down the rest of the signal chain for channel 1. If nothing is connected to the insert ports, it is bypassed and the signal is fed directly through the internal signal chain for that input channel. When you connect a cable to the insert output, the signal is almost always automatically rerouted away from the channel strip. You’ll need to feed something back into the insert input in order to continue using that channel strip on the mixing console.

There are two different connection designs for inserts on a mixing console. The ideal design is to have a separate ¼" or XLR connection for both the insert output and input. This allows you to use standard patch cables to connect the external processing equipment, and may also employ a balanced audio signal. If the company making the mixing console needs to save space or cut down on the cost of the console, they might decide to integrate both the insert output and input on a single ¼" TRS connector. In this case, the input and output are handled as unbalanced signals using the tip for one signal, the ring for the other signal, and a shared neutral on the sleeve. There is no standard for whether the input or output is carried on the tip vs. the ring. To use this kind of insert requires a special cable. This cable has three connectors. On one end is a ¼" TRS connector. This connector has two cables coming out of the end. One cable feeds an XLR male or a ¼" TS connector for the insert output and a XLR female or a ¼" TS connector for the insert input. Equalizer Section

After the gain section of the channel strip, the next section your audio signal encounters is the equalizer section (EQ) shown in Figure 7.25. The number of controls you see in this section of the channel strip varies greatly across the various models of mixing consoles. Very basic consoles may not include an EQ section at all. Generally speaking, the more money you pay for the console, the more knobs and buttons you find in the EQ section. We discussed the equalization process in depth in Chapter 7.

Figure 7.25 EQ section of an input channel strip

Figure 7.25 EQ section of an input channel strip on the Soundcraft K2 analog mixing console

Even the simplest of mixing consoles typically has two channels of EQ in each channel strip. These are usually a high shelf and a low shelf filter. These simple EQ sections consist of two knobs. One controls the gain for the high shelf and the other for the low shelf. The shelving frequency is a fixed value. If you pay a little more for your mixing console, you can get a third filter – a mid-frequency peak-notch filter. Again, the single knob is a gain knob with a fixed center frequency and bandwidth.

The next controllable parameter you’ll get with a nicer console is a frequency knob. Sometimes only the mid-frequency notch filter gets the extra variable center frequency knob, but the high and low shelf filters may get a variable filter frequency using a second knob as well. With this additional control, you now have a semi-parametric filter. If you are given a third knob to control the filter Q or Bandwidth, the filter becomes fully parametric. From there you simply get more bands of fully parametric filters per channel strip as the cost of the console increases.

Depending on your needs, you may not require five bands of EQ per channel strip. The option that is absolutely worth paying for is an EQ bypass button. This button routes the audio signal in the channel around the EQ circuit. This way, the audio signal doesn’t have to be processed by the EQ if you don’t need any adjustments to the frequency response of the signal. Routing around the EQ solves two potential problems. The first is the problem of inheriting someone else’s solution. There are a lot of knobs on a mixing console, and they aren’t always reset when you start working on a new project. If the EQ settings from a previous project are still dialed in, you could be inheriting a frequency adjustment that's not appropriate for your project. Having an EQ bypass button is a quick way to turn off all the EQ circuits so you're starting with a clean slate. The bypass button can also help you quickly do an A/B comparison without having to readjust all of the filter controls. The second problem is related to noise floor. Even if you have all the EQ gain knobs flattened out (no boost or cut), your signal is still passing though all those circuits and potentially collecting some noise along the way. Bypassing the EQ allows you to avoid that unnecessary noise. Auxiliaries

The Auxiliary controls in the channel strip are shown in Figure 7.26. Each auxiliary send knob represents an additional physical audio path/output on the mixing console. As you increase the value of an auxiliary send knob, you're setting a certain level of that channel’s signal to be sent into that auxiliary bus. As each channel is added into the bus to some degree, a mix of those sounds is created and sent to a physical audio output connected to that bus. You can liken the function of the auxiliary busses to an actual bus transportation system. Each bus, or bus line, travels to a unique destination, and the send knob controls how much of that signal is getting on the bus to go there. In most cases, the mixing console will also have a master volume control to further adjust the combined signal for each auxiliary output. This master control can be a fader or a knob and is usually located in the central control section of the mixing console.

An auxiliary is used whenever you need to send a unique mix of the various audio signals in the console to a specific device or person. For example, when you record a band, the lead singer wears headphones to hear the rest of the band as well as her own voice. Perhaps the guitar is the most important instrument for the singer to hear because the guitar contains the information about the right pitch the singer needs to use with her voice. In this situation, you would connect her headphones to a cable that is fed from an auxiliary output, which we'll call “Aux 1,” on the mixing console. You might dial in a bit of sound to Aux 1 across each input channel of the mixing console, but on the channels containing the guitar and the singer’s own vocals the Aux 1 controls would be set to a higher value so they're louder in the mix being sent to the singer’s headphones.

Figure 7.26 Auxiliary section of input channel strip

Figure 7.26 Auxiliary section of input channel strip on the Soundcraft K2 mixing console

The auxiliary send knobs on an input channel strip come in two configurations. Pre-Fader aux sends a signal level into the aux bus independently of the position of the channel fader. In our example of the singer in the band, a pre-fade aux would be desirable because once you've dialed in an aux mix that works for the singer, you don’t want that mix changing every time you adjust the channel fader. When you adjust the channel fader, it's in response to the main mix that is heard in the control room, which has no bearing on what the singer needs to hear.

The other configuration for an aux send is Post-Fader. In this case, dialing in the level on the aux send knob represents a level relative to the fader position for that input channel. So when the main mix is changed via the fader, the level in that aux send is changed as well. This is particularly useful when you're using an aux bus for some kind of effect processing. In our same recording session example, you might want to add some reverberation to the mix. Instead of inserting a separate reverb processor on each input channel, requiring multiple processors, it's much simpler to connect an aux output on the mixing console to the input of a single reverb processor. The output of the reverb processor then comes back into an unused input channel on the mixing console. This way, you can use the aux sends to dial in the desired amount of reverb for each input channel. The reverb processor then returns a reverberant mix of all of the sounds that gets added into the main mix. Once you get a good balance of reverb dialed in on an aux send for a particular input channel, you don’t want that balance to change. If the aux send to the reverb is pre-fader, when the fader is used to adjust the channel level within the main mix, the reverb level remains the same, disrupting the balance you achieve. Instead, when you turn up or down the channel fader, the level of the reverb should also increase or decrease respectively so the balance between the dry and the reverberant (wet) sound stays consistent. Using a post-fader aux send accomplishes this goal.

Some mixing consoles give you a switch to change the behavior of an aux bus between pre-fader and post-fader, while in other consoles this behavior may be fixed. Sometimes this switch is located next to the aux master volume control, and changes the pre-fader or post-fader mode for all of the channel aux sends that feed into that bus. More expensive consoles allow you to select pre- or post-fader behavior in a channel-specific way. In other words, each individual aux send dial on an input channel strip has its own pre- or post-fade button. With this flexibility Aux 1 can be set as a pre-fade aux for input channel 1 and a post-fade aux for input channel 2. Fader and Routing Section

The fader and routing section shown in Figure 7.27 is where you usually spend most of your time working with the console in an iterative fashion during the artistic process of mixing. The fader is a vertical slider control that adjusts the level of the audio signal sent to the various mixes you've routed on that channel. There are two common fader lengths: 60 mm and 100 mm. The 100 mm faders give your fingers greater range and control and are easier to work with. The fader is primarily an attenuator. It reduces the level of the signal on the channel. Once you've set the optimal level for the incoming signal with the preamplifier, you use the fader to reduce that level to something that fits well in the mix with the other sounds. The fader is a very low-noise circuit, so you can really set it to any level without having adverse effects on signal-to-noise ratio. One way to think about it is that the preamplifier is where the science happens; the fader is where the art happens. The fader can reduce the signal level all the way to nothing (−∞ or –inf), but typically has only five to ten dB on the amplification end of the level adjustment scale. When the fader is set to 0 dB, also referred to as unity, the audio signal passes through with no change in level. You should set the fader level to whatever sounds best, and don’t be afraid to move it around as the levels change over time.

Figure 7.27 Fader and routing section of an input channel strip

Figure 7.27 Fader and routing section of an input channel strip on the Soundcraft K2 analog mixing console

Near the fader there is usually a set of signal routing buttons. These buttons route the audio signal at a fixed level relative to the fader position to various output channels on the mixing console. There is almost always a main left and right stereo output (labeled “MIX” in Figure 7.27), and sometimes a mono or center output. Additionally, you may also be able to route the signal to one or more group outputs or subgroup mixes. A subgroup (sometimes, as with auxiliaries, also called a bus) represents a mixing channel where input signals can be grouped together under a master volume control before being passed on to the main stereo or mono output, as shown in Figure 7.28. An example of subgroup routing would be to route all the drum microphones to a subgroup so you can mix the overall level of the drums in the main mix using only one fader. A group is essentially the same thing, except it also has a dedicated physical output channel on the mixing console. The terms bus, group, and subgroup are often used interchangeably. Group busses are almost always post fader, and unlike auxiliary busses don't have variable sends – it’s all or nothing. Group routing buttons are often linked in stereo pairs, where you can use the pan knob to pan the signal between the paired groups, in addition to panning between the main stereo left and right bus.

Figure 7.28 Master control section of an analog mixing console

Figure 7.28 Master control section of the Soundcraft K2 analog mixing console

Also near the fader you usually have a mute button. The mute button mimics the behavior of pulling the input fader all the way down. In this case, pre-fade auxiliaries would continue to function. The mute button comes in handy when you want to stop hearing a particular signal in the main mix, but you don’t want to lose the level you have set on the fader or lose any auxiliary functionality, like signal being sent a headphone or monitor mix. Instead of a mute button, you may see an on/off button. This button shuts down the entire channel strip. In that situation, all signals stop on the channel, including groups, auxiliaries, and direct outs. Just to confuse you, manufacturers may use the terms mute and on/off interchangeably so in some cases, a mute button may behave like an on/off button and vice versa. Check the user manual for the mixing console to find out the exact function of your button.

Next to the fader there is typically a pre-fade listen (PFL) or a solo button. Pressing the PFL button routes the signal in that channel strip to a set of headphones or studio monitor outputs. Since it is pre-fade, you can hear the signal in your headphones even if the fader is down or the mute button is pressed. This is useful when you want to preview the sound on that channel before you allow it to be heard via your main or group outputs. If you have a solo button, when pressed it will also mute all the other channels, allowing you to hear only the solo-enabled channels. Solo is found in recording studio consoles or audio recording software. Sometimes the terms PFL and solo are used interchangeably so, again, check the user manual for your mixing console to be sure of the function for this button.

Similar to PFL is after-fade listen (AFL). AFL is found on output faders allowing you to preview in your headphones the signal that is passing through a subgroup, group, aux, or main output. The after-fade feature is important because it allows you to hear exactly what is passing through the output, including the level of the fader. For example, if a musician says that he can’t hear a given instrument in his monitor, you can use the AFL feature for the aux that feeds that monitor to see if the instrument can be heard. If you can hear it in your headphones, then you know that the aux is functioning properly. In this case, you may need to adjust the mix in that aux to allow the desired instrument to be heard more easily. If you can't hear the desired instrument in your headphones, then you know that you have a routing problem in the mixing console that's preventing the signal from sending out from that aux output.

Depending on the type of mixing console you're using, you may also have some sort of PPM (Peak Program Meter) near the fader. In some cases, this will be at the top of the console on a meter bridge. Cheaper consoles will just give you two LED indicators, one for when audio signal is present and another for when the signal clips. More expensive consoles will give you high-resolution PPMs with several different level indicators. A PPM is more commonly found in digital systems, but is also used in analog equipment. A PPM is typically a long column of several LED indicators in three different colors, as shown in Figure 7.29. One color represents signal levels below the nominal operating level, another color represents signals at or above nominal level, and the third color (usually red) represents a signal that is clipping or very near to clipping. A PPM responds very quickly to the audio signal. Therefore, a PPM is very useful for measuring peak values in an audio signal. If you’re trying to find the right position for a preamplifier, a PPM will show you exactly when the signal clips. Most meters in audio software are programmed to behave like a PPM.

Figure 7.29 PPM meters on a hardware mixing console and in a software DAW

Figure 7.29 PPM meters on a hardware mixing console and in a software DAW

7.2.2 Applying EQ

An equalizer can be incredibly useful when used appropriately, and incredibly dangerous when used inappropriately. Knowing when to use an EQ is just as important and knowing how to use it to accomplish the effect you are looking for. Every time you think you want to use an EQ you should evaluate the situation against this rule of thumb: EQ should be used to create an effect, not to solve a problem. Using an EQ as a problem solver can cause new problems when you should really just figure out what’s causing the original problem and fix that instead. Only if the problem can’t be solved in any other way should you pull up the EQ – for example, if you’re working post-production on a recording captured earlier during a film shoot, or you’ve run into an acoustical issue in a space that can’t be treated or physically modified. Rather than solving problems, you should try to use an EQ as a tool to achieve a certain kind of sound. Do you like your music to be heavy on the bass? An EQ can help you achieve this. Do you really like to hear the shimmer of the cymbals in a drum set? An EQ can help.

Let’s examine some common problems you may encounter where you might be tempted to use an EQ inappropriately. As you listen to the recording you’re making of a singer you notice that the recorded audio has a lot more low frequency than high frequency content, leading to a decreased intelligibility. You go over and stand next to the performer to hear what he actually sounds like and notice that he sounds quite different than what you’re hearing from the microphone. Standing next to him, you can hear all those high frequencies quite well. In this situation you may be tempted to pull out your EQ and insert a high shelf filter to boost all those high frequencies. This should be your last resort. Instead, you might notice that the singer is singing into the side of the microphone instead of the front. Because microphones are more directional at high frequencies than low frequencies, singing into the side of the microphone would mean that the microphone picks up the low frequency content very easily but the high frequencies are not being captured very well. In this case you would be using an EQ to boost something that isn’t being picked up very well in the first place. You will get much better results by simply rotating the microphone so it is pointed directly at the singer so the singer is singing into the part of the microphone that is more sensitive to high frequencies.

Another situation you may encounter would be when mixing the sound from multiple microphones either for a live performance or a recording. You notice as you start mixing everything together that a certain instrument has a huge dip around 250 Hz. You might be tempted to use an EQ to increase 250 Hz. The important thing to keep in mind here is that most microphones are able to pick up 250 Hz quite well from every direction, and it is unlikely that the instrument itself is somehow not generating the frequencies in the 250 Hz range while still generating all the other frequencies reasonably well. So before you turn on that EQ, you should mute all the other channels on the mixer and listen to the instrument alone. If the problem goes away, you know that whatever is causing the problem has nothing to do with EQ. In this situation, comb filtering is the likely culprit. There’s another microphone in your mix that was nearby and happened to be picking up this same instrument at a slightly longer distance of about two feet. When you mix these two microphones together, 250 Hz is one of the frequencies that cancels out. If comb filtering is the issue, you should try to better isolate the signals either by moving the microphones farther apart or preventing them from being summed together in the mix. A gate might come in handy here, too. If you gate both signals you can minimize the times when both microphones are mixed together, since the signals won’t be let through when the instruments they are being used for aren’t actually playing.

If comb filtering isn’t the issue, try moving a foot or two closer to or farther away from the loudspeakers. If the 250 Hz dip goes away in this case, there’s likely a standing wave resonance in your studio at the mix position that is cancelling out this frequency. Using an EQ in this case will not solve the problem since you’re trying to boost something that is actively being cancelled out. A better solution for the standing wave would be to consider rearranging your room or applying acoustical treatment to the offending surfaces that are causing this reflective build up.

Suppose you are operating a sound reinforcement system for a live performance and you start getting feedback through the sound system. When you hear that single frequency start its endless loop through the system you might be tempted to use an EQ to pull that frequency out of the mix. This will certainly stop the feedback, but all you really get is the ability to turn the system up another decibel or so before another frequency will inevitably start to feed back. Repeat the process a few times and in no time at all you will have completely obliterated the frequency response of your sound system. You won’t have feedback, but the entire system will sound horrible. A better strategy for solving this problem would be to get the microphone closer to the performer, and move the performer and the microphone farther away from the loudspeakers. You’ll get more gain this way and you can maintain the frequency response of your system. (Chapter 4 has moree on potential acoustic gain. Chapter 8 has an exercise on gain setting.)

We could examine many more examples of an inappropriate use of an EQ but they all go back to the rule of thumb regarding the use of an EQ as a problem solver. In most cases, an EQ is a very ineffective problem solver. It is, however, a very effective tool for shaping the tonal quality of a sound. This is an artistic effect that has little to do with problems of a given sound recording or reinforcement system. Instead you are using the EQ to satisfy a certain tonal preference for the listener. These effects could be as subtle as reducing an octave band of frequencies around 500 Hz by -3 dB to achieve more intelligibility for the human voice by allowing the higher frequencies to be more prominent. The effect could be as dramatic as using a bandpass filter to mimic the effect of a small cheap loudspeaker in a speakerphone. When using an EQ as an effect, keep in mind another rule of thumb. When using an EQ, you should reduce the frequencies that are too loud instead of increasing the frequencies that are too quiet. Every sound system, whether in a recording studio or a live performance, has an amplitude ceiling – the point at which the system clips and distorts. If you’ve done your job right, you will be running the sound system at an optimal gain, and a 3 dB boost of a given frequency on an EQ could be enough to cause a clipped signal. Reducing frequencies is always safer than boosting them since reducing them will not blow the gain structure in your signal path.

7.2.3 Applying Reverb

Aside:  In an attempt to reconcile these two schools of thought on reverberation in the recording studio, some have resorted to installing active acoustic systems in the recording studio. These systems involve placing microphones throughout the room that feed into live digital signal processors that generate thousands of delayed sounds that are then sent into several loudspeakers throughout the room. This creates a natural-sounding artificial reverb that is captured in the recording the same as natural reverb. The advantage here is that you can change the reverb by adjusting the parameters of the DSP for different recording situations. To hear an example of this kind of system in action, see this video from TRI Studios where Bob Weir from the Grateful Dead has installed an active acoustic system in his recording studio.

Almost every audio project you do will likely benefit from some reverb processing. In a practical sense, most of the isolation strategies we use when recording sounds will have a side effect of stripping the sound of natural reverberation. So anything recorded in a controlled environment such as a recording studio will probably need some reverb added to make it sound more natural. There are varying opinions on this among audio professionals. Some argue that artificial reverberation processers are sounding quite good now, and since it is impossible to remove natural reverberation from a recording, it makes more sense to capture your recorded audio as dry as possible. This way you’re able to artificially add back whatever reverberation you need in a way that you can control. Others argue that having musicians perform in an acoustically dry and isolated environment will negatively impact the quality of their performance. Think about how much more confident you feel when singing in the shower. All that reverberation from the tiled surfaces in the shower create a natural reverberation that makes your voice sound better to you than normal. That gives you the confidence to sing in a way that you probably don’t in public. So some recording engineers would prefer to have some natural reverberation in the recording room to help the musicians to deliver a better performance. If that natural reverberation is well controlled acoustically you could even end up with a recording that sounds pretty good already and might require minimal additional processing.

Regardless of the amount of reverb you already have in your recording, you will likely still want to add some artificial reverb to the mix. There are three places you can apply the reverb in your signal chain. You can set it up as an insert for a specific channel in a multi-channel mix. In this case the reverb only gets applied to the one specific channel, and the other channels are left unchanged. You have to adjust the wet/dry mix in the reverb processor to create an appropriate balance. This technique can be useful for a special effect you want to put on a specific sound, but using this technique on every channel in a large multi-channel mix costs you a lot in CPU performance because the multiple reverb processors that are running simultaneously. If you have a different reverb setting on each channel you could also have a rather confusing mix since every sound will seem to be in a different acoustic environment. Maybe that’s what you want if you’re creating a dream sequence or something abstract for a play or film, but for a music recording it usually makes more sense to have every instrument sounding like it is in the same room.

The second reverb technique can solve both the problem of CPU performance and varying acoustic signatures. In this case you would set up a mix bus that has a reverb inserted. You would set the reverb processor to 100% wet. This basically becomes the sound of your virtual room. Then you can set up each individual channel in your mix to have a variable aux send that dials in a certain amount of the signal into the reverb bus. In other words, the individual sends decide how much that instrument interacts with your virtual room. The individual channel will deliver the dry sound to the mix and the reverb bus will deliver the wet. The amount of sound that is sent on the variable aux send determines the balance of wet to dry. This strategy allows you to send many different signals into the reverb processor at different levels and therefore have a separate wet/dry balance for each signal, while using only one reverberation processor. The overall wet mix can also be easily adjusted using the fader on the aux reverb bus channel. This technique is illustrated in Figure 7.30.

Figure 7.30 Routing each channel through a single reverb bus

Figure 7.30 Routing each channel through a single reverb bus

The third strategy for applying reverberation is to simply apply a single reverb process to an entire mix output. This technique is usually not preferred because you have no control over the reverb balance between the different sounds in the mix. The reason you would use this technique is if you don’t have access to the raw tracks or if you are trying to apply a special reverb effect to a single audio file. In this case just pick a reverb setting and adjust the wet/dry mix until you achieve the sound you are looking for.

The most difficult task in using reverb is to find the right balance. It is very easy to overdo the effect. The sound of reverberation is so intoxicating that you have to constantly fight the urge to apply the effect more dramatically. Before you commit to any reverb effect, listen to it though a few different speakers or headphones and in a few different listening environments. A reverb effect sounds like a good balance in one environment might sound over the top in another. Listen to other mixes of similar music or sound to compare your work with the work of seasoned professionals. Before long you’ll develop a sixth sense for the kind of reverb to apply in a given situation.

7.2.4 Applying Dynamics Processing

When deciding whether to use dynamics processing you should keep in mind that a dynamics processor is simply an automatic volume knob. Any time you find yourself constantly adjusting the level of a sound, you may want to consider using some sort of dynamics processor to handle that for you. Most dynamics processors are in the form of downwards compressors. These compressors work by reducing the level of sounds that are too loud but letting quieter sounds pass without any change in level.

Aside:  There is some disagreement among audio professionals about the use of compressors. There are some who consider using a compressor as a form of cheating. Their argument is that no compressor can match the level of artistry that can be accomplished by a skilled mixer with their fingers on the faders. In fact, if you ask some audio mix engineers which compressors they use, they will respond by saying that they have ten compressors and will show them to you by holding up both hands and wiggling their fingers!

One example when compression can be helpful is when mixing multiple sounds together from a multitrack recording. The human voice singing with other instruments is usually a much more dynamic sound than the other instruments. Guitars and basses, for example are not known as particularly dynamic instruments. A singer is constantly changing volume throughout a song. This is one of the tools a singer uses to produce an interesting performance. When mixing a singer along with the instruments from a band, the band essentially creates a fairly stable noise floor. The word noise is not used here in a negative context; rather, it is used to describe a sound that is different from the vocal that has the potential of masking the vocal if there is not enough difference in level between the two. As a rule of thumb, for adequate intelligibility of the human voice, the peaks of the voice signal need to be approximately 25 dB louder than the noise floor, which in this case is the band. It is quite possible for a singer to perform with a 30 dB dynamic range. In other words, the quietest parts of the vocal performance are 30 dB quieter than the loudest parts of the vocal performance.

If the level of the band is more or less static and the voice is moving all around, how are you going to maintain that 25 dB ratio between the peaks of the voice and the level of the band? In this situation you will never find a single level for the vocal fader that will allow it to be heard and understood consistently throughout the song. You could painstakingly draw in a volume automation curve in your DAW software, or you could use a compressor to do it for you. If you can set the threshold somewhere in the middle of the dynamic range of the vocal signal and use a 2:1 or 4:1 compression ratio, can easily turn that 30 dB of dynamic range into a 20 dB range or less. Since the compressor is turning down all the loud parts, the compressed signal will sound much quieter than the uncompressed signal, but if you turn the signal up using either the output gain of the compressor or the channel fader you can bring it back to a better level. With the compressed signal, you can now much more easily find a level for the voice that allows it to sit well in the mix. Depending on how aggressive you are about the compression, you may still need to automate a few volume changes, but the compressor has helped turn a very difficult to solve problem into something more manageable.

Rather than using a compressor to allow a sound to more easily take focus over a background sound, you can also use compression as a tool for getting a sound to sit in the mix in a way that allows other sounds to take focus. This technique is used often in theatre and film for background music and sound effects. The common scenario is when a sound designer or composer tries to put in some underscore music or background sounds into a scene for a play or a film and the director inevitably says, “turn it down, it’s too loud.” You turn it down by 6 dB or so and the director still thinks it’s too loud. By the time you turn it down enough to satisfy the director, you can hardly hear the sound and before long, you’ll be told to simply cut it because it isn’t contributing to the scene in any meaningful way.

The secret to solving this problem is often compression. When the director says the sound is too loud, what he really means is that the sound is too interesting. More interesting than the actor, in fact, and consequently the audience is more likely to pay attention to the music or the background sound than they are to the actor. One common culprit when a sound is distracting is that it is too dynamic. If the music is constantly jumping up and down in level, it will draw your focus. Using a compressor to make the underscore music or background sounds less dynamic allows them to sit in the mix and enhance the scene without distracting from the performance of the actor.

Compression can be a useful tool, but like any good thing, if it’s overused compression can be detrimental to the quality of your sound. Dynamics are one quality of sound and music that makes it exciting, interesting, and evocative. A song with dynamics that have been completely squashed will not be very interesting to listen to and can cause great fatigue on the ears. Also, if you apply compression inappropriately, it may cause audible artifacts in the sound, where you can hear when the sound is being attenuated and released. This is referred to as “pumping” or “breathing,” and it usually means you’ve taken the compression too far or in the wrong direction. So you have to be very strategic about the use of compression and go easy on the compression ratio. Often, a mild compression ratio is enough to tame an overly dynamic sound without completely stripping it of all its character.

7.2.5 Applying Special Effects

One of the most common special effects is using delay to create an echo effect. This is used often in popular music. The challenge with a delay effect is to synchronize the timing of the echoes with the beat of the music. If you’re using a delay plug-in with a DAW program, the plug-in will try to use the metronome of your project file to create the delay timing. This works if you recorded the music to the system’s metronome, but if you just recorded everything freestyle you have to synchronize the delay manually. Typically this is done with a tap pad. Also called tap-delay, this plug-in uses a pad or button that you can tap along with the beat of the music to keep the echoes synchronized. Usually after eight taps, the echoes get in sync with the music, but as the performance from the musician changes, you need to periodically re-tap the plug-in. Figure 7.31 shows a tap delay processor with the mouse pointer on the tap pad.

Figure 7.31 A tap delay plug-in

Figure 7.31 A tap delay plug-in

Other special effects including flangers, pitch shifting/autotune, etc. may be applied in several different situations. There are really no rules with special effects. Just make sure you have a real reason for using the effect and don’t overdo it.

Section 7.3 - Science, Mathematics, and Algorithms

7.3.1 Convolution and Time Domain Filtering

In earlier chapters, we showed how audio signals can be represented in either the time domain or the frequency domain. In this section, you'll see how mathematical operations are applied in these domains to implement filters, delays, reverberation, etc. Let's start with the time domain.

Filtering in the time domain is done by a convolution operation. Convolution uses a convolution filter, whichis an array of N values that, when graphed, takes the basic shape shown in Figure 7.32. A convolution filter is also referred to as a convolution mask, an impulse response (IR), or a convolution kernel. There are two commonly-used time-domain convolution filters that are applied to digital audio. They are FIR filters (finite impulse response) and IIR filters (infinite impulse response).

Figure 7.32 Graph of time-domain convolution filter

Figure 7.32 Graph of time-domain convolution filter

Equation 7.1 describes FIR filtering mathematically.

\begin{matrix}\mathbf{y}\left ( n \right )=\mathbf{h}\left ( n \right )\otimes \mathbf{x}\left ( n \right )=\sum_{k=0}^{N-1}\mathbf{h}\left ( k \right )\mathbf{x}\left ( n-k \right )=\sum_{k=0}^{N-1}\mathbf{a}_{k}\mathbf{x}\left ( n-k \right )\\\textit{where N is the size of the filter and}\: x\left ( n-k \right )=0\: if\; n-k< 0\end{matrix}

Equation 7.1 FIR filter


You can actually think of equations such as Equation 7.1 in one of two ways: x  and y could be understood as vectors of audio samples in the time domain, being the input and being the output. The equation must be executed N times to get each nth output sample. Alternatively,  could be understood as a function with input n. The function is executed N times to yield all N output samples.

By our convention, boldface variables refer to vectors (i.e., arrays). In this equation, \mathbf{h}=\left [ a_{0},a_{1},a_{2}\cdots,a_{N-1} \right ] is the convolution filter – which is a vector of multipliers to be applied successively to audio samples. The number of multipliers is the order of a filter, N in this case. N is sometimes also referred to as the number of taps in the filter.

It's helpful to think of Equation 7.1 algorithmically, as described Algorithm 7.1. The notation \mathbf{y}\left ( n \right ) indicates that the n^{th} output sample is created from a convolution of input values from the audio signal x and the filter multipliers in h, as given in the summation. The equation is repeated in a loop for every sample in the audio input.


x, an array of digitized audio samples (i.e., in the time domain) of size M

h, a convolution filter of size N (Specifically, a finite-impulse-response filter, FIR


y, the audio samples, filtered



for \left ( n=0\: to\; N-1 \right ) {

\mathbf{y}\left ( n \right )=\mathbf{h}\left ( n \right )\otimes \mathbf{x}\left ( n \right )=\sum_{k=0}^{N-1}\mathbf{h}\left ( k \right )\mathbf{x}\left ( n-k \right )

where x\left ( n-k \right )=0\: if\; n-k< 0


Algorithm 7.1 Convolution with a finite impulse response (FIR) filter

The FIR convolution process is described diagrammatically in Figure 7.33.


Figure 7.33 Filtering in the time domain by convolving with an FIR filter

Figure 7.33 Filtering in the time domain by convolving with an FIR filter

You can implement convolution yourself as a function, or you can use MATLAB’s conv function. The conv function requires two arguments. The first is an array containing the coefficients of the filter, and the second is the array of time domain audio samples to be convolved. Say that we have an FIR filter of size 3 defined as

y\left ( n \right )=0.4\left ( n \right )+0.3x\left ( n-1 \right )+0.3x(n-2)

Equation 7.2 Equation for an example FIR filter

In MATLAB, the coefficients of the filter can be stored in an array h like this:

h = [0.4 0.3 0.3];


Then we can read in a sound file and convolve it and listen to the results.

x = wavread('ToccataAndFugue.wav');

y = conv(h,x);

sound(y, 44100);


(This is just an arbitrary example with a random filter to demonstrate how you convolve in MATLAB. We haven’t yet considered how the filter itself is constructed.)

IIR filters are also time domain filters, but the process by which they work is a little different. An IIR filter has an infinite length, given by this equation:

\begin{matrix}\mathbf{y}\left ( n \right )=\mathbf{h}\left ( n \right )\otimes \mathbf{x}\left ( n \right )=\sum_{k=0}^{\infty }\mathbf{h}\left ( k \right )\mathbf{x}\left ( n-k \right )\\where\; x\left ( n-k \right )=0\: if\; n-k< 0\\\textit{and k is theoretically infinite}\end{matrix}

Equation 7.3 IIR Filter, infinite form

We can't deal with an infinite summation in practice, but Equation 7.3 can be transformed to a difference equation form which gives us something we can work with.

\begin{matrix}\mathbf{y}\left ( n \right )=\mathbf{h}\left ( n \right )\otimes \mathbf{x}\left ( n \right )=\sum_{k=0}^{N-1}\mathbf{a}_{k}\mathbf{x}\left ( n-k \right )-\sum_{k=1}^{M}\mathbf{b}_{k}\mathbf{y}\left ( n-k \right )\\\textit{where N is the size of the forward filter, M is the size of the feedback filter, and}\; x\left ( n-k \right )=0\: if\; n-k< 0\end{matrix}

Equation 7.4 IIR filter, difference equation form

In Equation 7.4, N is the size (also called order) of the forward filter and M is the size of the feedback filter. The output from an IIR filter is determined by convolving the input and combining it with the feedback of previous output. In contrast, the output from an FIR filter is determined solely by convolving the input.

FIR and IIR filters each have their advantages and disadvantages. In general, FIR filters require more memory and processor time. IIR filters can more efficiently create a sharp cutoff between frequencies that are filtered out and those that are not. An FIR filter requires a larger filter size to accomplish the same sharp cutoff as an IIR filter. IIR filters also have the advantage of having analog equivalents, which facilitates their design. An advantage of FIR filters is that they can be constrained to have linear phase response, which means that phase shifts for frequency components are proportional to the frequencies. This is good for filtering music because harmonic frequencies are phase-shifted by the same proportions, preserving their harmonic relationship. Another advantage of FIR filters is that they're not as sensitive to the noise that results from low bit depth and round-off error.

The exercise associated with this section shows you how to make a convolution filter by recording an impulse response in an acoustical space. Also, in Section 7.3.9, we’ll show a way to apply an IIR filter in MATLAB.

7.3.2 Low-Pass, High-Pass, Bandpass, and Bandstop Filters

You may have noticed that in our discussion of frequency domain and time domain filters, we didn't mention how we got the filters – we just had them and applied them. In the case of an FIR filter, the filter is represented in the coefficients \left [ a_{0},a_{1},a_{2},\cdots ,a_{N-1} \right ]. In the case of the IIR filter, the filter resides in coefficients \left [ a_{0},a_{1},a_{2},\cdots ,a_{N-1} \right ] and \left [ b_{0},b_{1},b_{2},\cdots ,b_{M} \right ].

Without explaining in detail the mathematics of filter creation, we can show you algorithms for creating low-pass, high-pass, bandpass, and bandstop filters when they are given the appropriate parameters as input. Low-pass filters allow only frequencies below a cutoff frequency f_{c} to pass through. Thus, Algorithm 7.2 takes f_{c} as input and outputs an N-element array constituting a low-pass filter. Similarly, Algorithm 7.3 takes f_{c} as input and yields a high-pass filter, and Algorithm 7.4 and Algorithm 7.5 take f_{1} and f_{2} as input to yield bandpass and bandstop filters. These algorithms yield time-domain filters. If you're interested in how these algorithms were derived, see (Ifeachor and Jervis 1993), (Steiglitz 1996), or (Burg 2008).

algorithm FIR_low_pass filter



f_c, the cutoff frequency for the low-pass filter, in Hz

f_samp, sampling frequency of the audio signal to be filtered, in Hz

N, the order of the filter; assume N is odd


h, a low-pass FIR filter in the form of an N-element array */


//Normalize f_c and ω _c so that pi is equal to the Nyquist angular frequency

f_c = f_c/f_samp

ω_c = 2*pi*f_c

middle = N/2    /*Integer division, dropping remainder*/

for i = −N/2 to N/2

   if (i == 0) h(middle) = 2*f_c

   else h(i + middle) = sin(ω_c*i)/(pi*i)


Algorithm 7.2 Low-pass filter

algorithm FIR_high_pass filter



f_c, the cutoff frequency for the high pass filter, in Hz

f_samp, sampling frequency of the audio signal to be filtered, in Hz

N, the order of the filter; assume N is odd


h, a high-pass FIR filter in the form of an N-element array */


//Normalize f_c and ω _c so that pi is equal to the Nyquist angular frequency

f_c = f_c/f_samp

ω_c = 2*pi*f_c

middle = N/2    /*Integer division, dropping remainder*/

for i = −N/2 to N/2

   if (i == 0) h(middle) = 1 - 2*f_c

   else h(i + middle) = -sin(ω_c*i)/(pi*i)


Algorithm 7.3 High-pass filter

algorithm FIR_bandpass filter



f1, the lowest frequency to be included, in Hz

f2, the highest frequency to be included, in Hz

f_samp, sampling frequency of the audio signal to be filtered, in Hz

N, the order of the filter; assume N is odd


h, a bandpass FIR filter in the form of an N-element array */


//Normalize f_c and ω _c so that pi is equal to the Nyquist angular frequency

  f1_c = f1/f_samp

  f2_c = f2/f_samp

  ω1_c = 2*pi*f1_c

  ω2_c = 2*pi*f2_c

  middle = N/2    /*Integer division, dropping remainder*/

  for i = −N/2 to N/2

   if (i == 0) h(middle) = 2*(f2_c – f1_c)


  h(i + middle) = sin(ω2_c*i)/(pi*i) – sin(ω1_c*i)/(pi*i)


Algorithm 7.4 Bandpass filter

algorithm FIR_bandstop filter



f1, the highest frequency to be included in the bottom band, in Hz

f2, the lowest frequency to be included in the top band, in Hz

Everything from f1 to f2 will be filtered out

f_samp, sampling frequency of the audio signal to be filtered, in Hz

N, the order of the filter; assume N is odd


h, a bandstop FIR filter in the form of an N-element array */


//Normalize f_c and ω _c so that pi is equal to the Nyquist angular frequency

f1_c = f1/f_samp

f2_c = f2/f_samp

ω1_c = 2*pi*f1_c

ω2_c = 2*pi*f2_c

middle = N/2    /*Integer division, dropping remainder*/

for i = −N/2 to N/2

   if (i == 0) h(middle) = 1 - 2*(f2_c – f1_c)


h(i + middle) = sin(ω1_c*i)/(p*i) – sin(ω2_c*i)/(pi*i)


Algorithm 7.5 Bandstop filter

As an exercise, you can try implementing these algorithms in C++, Java, or MATLAB and see if they actually work. In Section 7.3.8, we'll show you some higher level tools in MATLAB's digital signal processing toolkit that create these types of filters for you.

7.3.3 The Z-Transform

If you continue to work in digital signal processing, you need to be familiar with the z-transform, a mathematical operation that converts a discrete time domain signal into a frequency domain signal represented as real or complex numbers. The formal definition is below.

Let \mathbf{x}\left ( n \right ) be a sequence of discrete values for n=0,1,\cdots . The one-sided z-transform of \mathbf{x}\left ( n \right ), \mathbf{X}\left ( z \right ), is a function that maps from complex numbers to complex numbers as follows:

\begin{matrix}\mathbf{X}\left ( z \right )=\sum_{n=0}^{\infty }\mathbf{x}\left ( n \right )z^{-n}\\\textit{where z is a complex number}\end{matrix}

Equation 7.5 The one-sided z-transform

A full z-transform sums from -∞ to ∞, but the one-sided transform suffices for us because n isan index into an array of audio samples and will never be negative.

In practice, our sector of time domain audio samples, has a finite length N. Thus, if we assume that for \mathbf{x}\left ( n \right )=0\; for\; n\geq N, then we can redefine the transform as

\mathbf{X}\left ( z \right )=\sum_{n=0}^{N-1}\mathbf{x}\left ( n \right )z^{-n}

\mathbf{X}\left ( z \right ) is a discrete function from complex variables to complex variables. Now, look at what results if we set z=e^{\frac{i2\pi k}{N}} and apply the z-transform to a vector of length N (applying the equation N times for 0\leq k< N). This yields the following:

for\; 0\leq k< N,\mathbf{X}\left ( Z_{k} \right )=\sum_{n=0}^{N-1}\mathbf{x}\left ( n \right )\left ( e^{i2\pi k} \right )^{-n}=\sum_{n=0}^{N-1}\mathbf{x}\left ( n \right )e^{\frac{-i2\pi kn}{N}}=\sum_{n=0}^{N-1}\mathbf{x}\left ( n \right )\cos ^{\frac{2\pi kn}{N}}-i\mathbf{x}\left ( n \right )\sin ^{\frac{2\pi kn}{N}}

Equation 7.6 The discrete Fourier transform as a special case of the z-transform

The idea of Equation 7.6 is that we are evaluating the equation for each k^{th} frequency component \left ( \frac{2\pi k}{N} \right ) to determine N frequency components in the audio data. You can see that this equation is equivalent to the definition of the discrete Fourier transform given in Chapter 2. The Fourier transform transforms from the time domain (real numbers) to the frequency domain (complex numbers), but since real numbers are a subset of the complex numbers, the Fourier transform is an instance of a z-transform.

7.3.4 Filtering in the Time and Frequency Domains

In Chapter 2, we showed how the Fourier transform can be applied to transform audio data from the time to the frequency domain. The Fourier transform is a function that maps from real numbers (audio samples in the time domain) to complex numbers (frequencies, which have magnitude and phase). We showed in the previous section that the Fourier transform is a special case of the z-transform, a function that maps from complex numbers to complex numbers. Now let’s consider the equivalence of filtering in the time and frequency domains. We begin with some standard notation.

Say that you have a convolution filter as follows:

\mathbf{y}\left ( n \right )=\mathbf{h}\left ( n \right )\otimes \mathbf{x}\left ( n \right )

By convention, the z-transform of \mathbf{x}\left ( n \right ) is called \mathbf{X}\left ( z \right ), the z-transform of \mathbf{y}\left ( n \right )is called \mathbf{Y}\left ( z \right ), and the z-transform of \mathbf{h}\left ( n \right ) is called \mathbf{H}\left ( z \right ), as shown in Table 7.1.

time domain frequency domain
input audio \mathbf{x}\left ( n \right ) \mathbf{X}\left ( z \right )
output, filtered audio \mathbf{y}\left ( n \right ) \mathbf{Y}\left ( z \right )
filter \mathbf{h}\left ( n \right ) \mathbf{H}\left ( z \right )
filter operation \mathbf{y}\left ( n \right )=\mathbf{h}\left ( n \right )\otimes \mathbf{x}\left ( n \right ) \mathbf{X}\left ( z \right )=\mathbf{H}\left ( z \right )\ast \mathbf{Y}\left ( z \right )

Table 7.1 Conventional notation for time and frequency domain filters

The convolution theorem is an important finding in digital signal processing that shows us the equivalence of filtering in the time domain vs. the frequency domain. By this theorem, instead of filtering by a convolution, as expressed in \mathbf{y}\left ( n \right )=\mathbf{h}\left ( n \right )\otimes \mathbf{x}\left ( n \right ), we can filter by multiplication, as expressed in \mathbf{Y}\left ( z \right )=\mathbf{H}\left ( z \right )\ast \mathbf{X}\left ( z \right ). (Note that \mathbf{H}\left ( z \right )\ast \mathbf{X}\left ( z \right ) is an element-by-element multiplication of order N.)  That is, if you take a time domain filter, transform it to the frequency domain, transform your audio data to the frequency domain, multiply the frequency domain filter and the frequency domain audio data, and do the inverse Fourier transform on the result, you get the same result as you would get by convolving the time domain filter with the time domain audio data. This is in essence the convolution theorem, explained diagrammatically in Figure 7.34. In fact, with a fast implementation of the Fourier transform, known as the Fast Fourier Transform (FFT), filtering in the frequency domain is more computationally efficient than filtering in the time domain. That is, it takes less time to do the operations.

Figure 7.34 The Convolution Theorem

Figure 7.34 The Convolution Theorem

MATLAB has a function called fft for performing the Fourier transform on a vector of audio data and a function called conv for doing convolutions. However, to get a closer view of these operations, it may be enlightening to try implementing these functions yourself. You can implement the Fourier transform either directly from Equation 7.6, or you can try the more efficient algorithm, the Fast Fourier transform. To check your accuracy, you can compare your results with MATLAB’s functions. You can also compare the run times of programs that filter in the time vs. the frequency domain. We leave this as an activity for you to try.

In Chapter 2, we discussed the necessity of applying the Fourier transform over relatively small segments of audio data at a time. These segments, called windows, are on the order of about 1024 to 4096 samples. If you use the Fast Fourier transform, the window size must be a power of 2. It’s not hard to get an intuitive understanding of why the window has to be relatively small. The purpose of the Fourier transform is to determine the frequency components of a segment of sound. Frequency components relate to pitches that we hear. In most sounds, these pitches change over time, so the frequency components change over time. If you do the Fourier transform on, say, five seconds of audio, you’ll get an imprecise view of the frequency components in that time window, called time slurring. However, what if you choose a very small window size, say just one sample? You couldn’t possibly determine any frequencies in one sample, which at a sampling rate of 44.1 kHz is just 1/44,100 second. Frequencies are determined by how a sound wave’s amplitude goes up and down as time passes, so some time must pass for there to be such a thing as frequency.

The upshot of this observation is that the discrete Fourier transform has to be applied over windows of samples where the windows are neither too large nor too small. Note that the window size has a direct relationship with the number of frequency components you detect. If your window has a size of N, then you get an output telling you the magnitudes of N/2 frequency bands from the discrete Fourier transform, ranging in frequency from 0 to the Nyquist frequency (i.e., ½ the sampling rate). An overly small window in the Fourier transform gives you very high time resolution, but tells you the magnitudes of only a small number of discrete, wide bands of frequencies. An overly large window yields many frequency bands, but with poor time resolution that leads to slurring. You want to have good enough time resolution to be able to reconstruct the resulting audio signal, but also enough frequency information to apply the filters with proper effect. Choosing the right window size is a balancing act.

7.3.5 Defining FIR and IIR Filters with Z-Transforms, Filter Diagrams, and Transfer Functions

Now let’s look at how the z-transform can be useful in describing, designing, and analyzing audio filters. We know that we can do filtering in the frequency domain with the operation

\mathbf{Y}\left ( z \right )=\mathbf{H}\left ( z \right )\ast \mathbf{X}\left ( z \right )

Equation 7.7 Filtering in the frequency domain

If we have the coefficients defining an FIR filter, \left [ a_{0},a_{1}.a_{2}\cdots ,a_{N-1} \right ], then \mathbf{H}\left ( z \right ) is derived directly from the definition of the z-transform

\mathbf{H}\left ( z \right )=\sum_{n=0}^{N-1}\mathbf{h}\left ( n \right )z^{-n}=a_{0}+a_{1}z^{-1}+a_{2}z^{-2}+\cdots +a^{N-1}z^{-\left ( N-1 \right )}

Equation 7.8 H(z), an FIR filter in the frequency domain

Consider our previous example of an FIR filter,

y\left ( n \right )=0.4x\left ( n \right )+0.3x\left ( n-1 \right )+0.3x\left ( n-2 \right )


where \mathbf{h}=\left [ a_{0},a_{1},a_{2} \right ]=\left [ 0.4,0.3,0.3 \right ]. This filter can be transformed to the frequency domain by applying the z-transform.

\mathbf{H}\left ( z \right )=\sum_{n=0}^{N-1}\mathbf{h}\left ( n \right )z^{-n}=a_{0}+a_{1}z^{-1}+a_{2}z^{-2}+\cdots +a^{N-1}z^{-\left ( N-1 \right )}=0.4+0.3z^{-1}+0.3z^{-2}

Filters can be expressed diagrammatically in terms of the z-transform, as illustrated in Figure 7.35. The z^{-m} elements are delay elements.

Figure 7.35 Example FIR filter diagram

Figure 7.35 Example FIR filter diagram

To derive \mathbf{H}\left ( z \right ) for an IIR filter, we rearrange Equation 7.7 as a ratio of the output to the input.

\mathbf{H}\left ( z \right )=\frac{\mathbf{Y}\left ( z \right )}{\mathbf{X}\left ( z \right )}

Equation 7.9 Filter in the frequency domain as a ratio of output to input

This form of the filter is called the transfer function. Using the definition of the IIR filter, \mathbf{y}\left ( n \right )=\mathbf{h}\left ( n \right )\otimes \mathbf{x}\left ( n \right )=\sum_{k=0}^{N-1}a_{k}\mathbf{x}\left ( n-k \right )-\sum_{k=1}^{M}b_{k}\mathbf{y}\left ( n-k \right ), we can express the IIR filter in terms of its coefficients \left [ a_{0},a_{1},a_{2}\cdots ,a_{N-1} \right ] and \left [ b_{1},b_{2},b_{3}\cdots ,b_{M} \right ], which yields the following transfer function defining the IIR filter:

\mathbf{H}\left ( z \right )=\frac{\mathbf{Y}\left ( z \right )}{\mathbf{X}\left ( z \right )}=\frac{a_{0}+a_{1}z^{-1}+a_{2}z^{-2}+\cdots +a_{N-1}z^{N-1}}{1+b_{1}z^{-1}+b_{2}z^{-2}+\cdots +b_{M}z^{N-M}}=\frac{a_{0}z^{N}+a_{1}z^{N-1}+\cdots +a_{N-1}z^{2N-1}}{z^{N}+b_{1}z^{N-1}+\cdots +b_{m}z^{2N-M}}

Equation 7.10 Transfer function for an IIR filter

(The usefulness of transfer functions will become apparent in the next section.) An IIR filter in the frequency domain can be represented with a filter diagram as shown in Figure 7.36.

Figure 7.36 Example IIR filter diagram

Figure 7.36 Example IIR filter diagram

7.3.6 Graphing Filters and Plotting Zero-Pole Diagrams

The function \mathbf{H}\left ( z \right ) defining a filter has a complex number z as its input parameter. Complex numbers are expressed as a+bi, where a is the real number component, b is the coefficient of the imaginary component, and i is \sqrt{-1}. The complex number plane can be graphed as shown in Figure 7.37, with the real component on the horizontal axis and the coefficient of the imaginary component on the vertical axis. Let’s look at how this works.

With regard to Figure 7.37, let x, y, and r denote the lengths of line segments \overline{AB}, \overline{BC}, and \overline{AC} , respectively. Since the circle is a unit circle, r=1. From trigonometric relations, we get

x=\cos \theta

y=\sin \theta

Point C represents the number \cos \theta +i\sin \theta on the complex number plane. By Euler’s identity, \cos \theta +i\sin \theta=e^{i\theta }.  \mathbf{H}\left ( z \right ) is a function that is evaluated around the unit circle and graphed on the complex number plane. Since we are working with discrete values, we are evaluating the filter’s response at N discrete points evenly distributed around the unit circle. Specifically, the response for the k^{th} frequency component is found by evaluating \mathbf{H}\left ( z_{k} \right ) at z=e^{i\theta }=e^{i\frac{2\pi k}{N}} where N is the number of frequency components. The output,  \mathbf{H}\left ( z_{k} \right ), is a complex number, but if you want to picture the output in the graph, you can consider only the magnitude of \mathbf{H}\left ( z_{k} \right ), which would be a real number coming out of the complex number plane (as we’ll show in Figure 7.41). This helps you to visualize how the filter attenuates or boosts the magnitudes of different frequency components.

Figure 7.37 A circle with radius 1 on the complex number plane

Figure 7.37 A circle with radius 1 on the complex number plane, with \theta =\frac{2\pi k}{N}

This leads us to a description of the zero-pole diagram, a graph of a filter’s transfer function that marks places where the filter does the most or least attenuation (the zeroes and poles, as shown in Figure 7.38). As the numerator of \frac{\mathbf{Y}\left ( z \right )}{\mathbf{X}\left ( z \right )} goes to 0, the output frequency is getting smaller relative to the input frequency. This is called a zero. As the denominator goes to 0, the output frequency is getting larger relative to the input frequency. This is called a pole.

To see the behavior of the filter, we can move around the unit circle looking at each frequency component at point e^{i\theta } where \theta =\frac{2\pi k}{N} for 0\leq k< N-1. Let d_{zero} be the point’s distance from a zero and d_{pole} be the point’s distance from a pole. If, as you move from point P_{k} to point P_{k+1} (these points representing input frequencies e^{i\frac{2\pi k}{N}} and e^{i\frac{2\pi\left ( k+1 \right )}{N}} respectively), the ratio \frac{d_{zero}}{d_{pole}} gets larger, then the filter is progressively allowing more of the frequency e^{i\frac{2\pi k}{N}} to “pass through” as opposed to e^{i\frac{2\pi\left ( k+1 \right )}{N}}.

Let’s try this mathematically with a couple of examples. These examples will get you started in your understanding of zero-pole diagrams and the graphing of filters, but the subject is complex. MATLAB’s Help has more examples and more explanations, and you can find more details in the books on digital signal processing listed in our references.

First let’s try a simple FIR filter. The coefficients for this filter are . In the time domain, this filter can be seen as a convolution represented as follows:

\mathbf{y}\left ( n \right )=\sum_{k=0}^{1}\mathbf{h}\left ( k \right )\mathbf{x}\left ( n-k \right )=\mathbf{x}\left ( n \right )-0.5\mathbf{x}\left ( n-1 \right )

Equation 7.11 Convolution for a simple filter in the time domain

Converting the filter to the frequency domain by the z-transform, we get

\mathbf{H}\left ( z \right )=1-0.5z^{-1}=\frac{z\left ( 1-0.5z^{-1} \right )}{z}=\frac{z-0.5}{z}

Equation 7.12 Delay filter represented as a transfer function

In the case of this simple filter, it’s easy to see where the zeroes and poles are. At z=0.5, the numerator is 0. Thus, a zero is at point (0.5, 0). At z = 0, the denominator is 0. Thus, a pole is at (0,0). There is no imaginary component to either the zero or the pole, so the coefficient of the imaginary part of the complex number is 0 in both cases.

Once we have the zeroes and poles, we can plot them on the graph representing the filter, \mathbf{H}\left ( z \right ), as shown in Figure 7.38. To analyze the filter’s effect on frequency components, consider the k^{th} discrete point on the unit circle, point P_{k}, and the behavior of filter \mathbf{H}\left ( z \right ) at that point. Point P_{0} is at the origin, and the angle formed between the x-axis and the line between P_{0} and P_{k}is \theta =\frac{2\pi k}{N}. Each such point P_{k} at angle \theta =\frac{2\pi k}{N} corresponds to a frequency component e^{i}\frac{2\pi k}{N} in the audio that is being filtered. Assume that the Nyquist frequency has been normalized to \pi. Thus, by the Nyquist theorem, the only valid frequency components to consider are between 0 and \pi. Let d_{zero} be the distance between P_{k} and a zero of the filter, and let d_{pole} be the distance between P_{k} and a pole. The magnitude of the frequency response, \mathbf{H}\left ( z \right ), will be smaller in proportion to \frac{d_{zero}}{d_{pole}}. For this example, the pole is at the origin so for all points P_{k}, d_{pole} is 1. Thus, the magnitude of \frac{d_{zero}}{d_{pole}} depends entirely on d_{zero}. As d_{zero} gets larger, the frequency response gets larger. d_{zero} increases as you move from \theta =0 to \theta =\pi. Since the frequency response increases as frequency increases, this filter is a high-pass filter, attenuating low frequencies more than high ones.

Figure 7.38 Zero-pole graph of simple delay filter

Figure 7.38 Zero-pole graph of simple delay filter

Now let’s try a simple IIR filter. Say that the equation for the filter is this:

\mathbf{y}\left ( n \right )=\mathbf{x}\left ( n \right )-\mathbf{x}\left ( n-2 \right )-0.49\mathbf{y}\left ( n-2 \right )

Equation 7.13 Equation for an IIR filter

The coefficients for the forward filter are \left [ 1,0,-1\right ], and the coefficients for the feedback filter are \left [ 0, -0.49\right ]. Applying the transfer function form of the filter, we get

\mathbf{H}\left ( z \right )=\frac{1-z^{-2}}{1-0.49z^{-2}}=\frac{z^{2}-1}{z^{2}-0.49}=\frac{\left ( z+1 \right )\left ( z-1 \right )}{\left ( z+0.7i \right )\left ( z-0.7i \right )}

Equation 7.14 Equation for IIR filter, version 2

From Equation 7.14, we can see that there are zeroes at z=1 and z=-1 and poles at z=0.7i and z=-0.7i. The zeroes and poles are plotted in Figure 7.39. This diagram is harder to analyze by inspection because there are two zeroes and two poles.

Figure 7.39 A zero-pole diagram for an IIR filter with two zeroes and two poles

Figure 7.39 A zero-pole diagram for an IIR filter with two zeroes and two poles

If we graph the frequency response of the filter, we can see that it’s a bandpass filter. We can do this with MATLAB’s fvtool (Filter Visualization Tool). The fvtool expects the coefficients of the transfer function (Equation 7.10) as arguments. The first argument of fvtool should be the coefficents of the numerator \left ( a_{0},a_{1},a_{2},\cdots \right ), and the second should be the coefficients of the denominator \left ( b_{1},b_{2},b_{3},\cdots \right ). Thus, the following MATLAB commands produce the graph of the frequency response, which shows that we have a bandpass filter with a rather wide band (Figure 7.40).

a = [1 0 -1];

b = [1 0 -0.49];



(You have to include the constant 1 as the first element of b.) MATLAB’s fvtool is discussed further in Section 7.3.9.

Figure 7.40 Frequency response of bandpass filter

Figure 7.40 Frequency response of bandpass filter

We can also graph the frequency response on the complex number plane, as we did with the zero-pole diagram, but this time we show the magnitude of the frequency response in the third dimension by means of a surface plot. This plot is shown in Figure 7.41. Writing a MATLAB function to create a graph such as this is left as an exercise.

Figure 7.41 Surface plot of filter on complex number plane, with zeroes marked in red

Figure 7.41 Surface plot of filter on complex number plane, with zeroes marked in red

7.3.7 Comb Filters and Delays

In Chapter 4, we discussed comb filtering in the context of acoustics and showed how comb filters can be made by means of delays. Let’s review comb filtering now.

A comb filter is created when two identical copies of a sound are summed (in the air, by means of computer processing, etc.), with one copy being offset in time from the other. The amount of offset determines which frequencies are combed out. The equation that predicts the combed frequency is given below.

\begin{matrix}\textrm{Given a delay of }t\textrm{ seconds between two identical copies of a sound, then the frequencies of }f_{i}\textrm{ that will be combed are }\\ f_{i}=\frac{2i+1}{2t}\textrm{ for all integers }i\geq 0\end{matrix}

Equation 7.15 Comb filtering

For a sampling rate of 44.1 kHz, 0.01 is equivalent to 441 samples. This delay is demonstrated in the MATLAB code below, which applies a delay of 0.01 seconds to white noise and graphs the resulting frequencies. (The white noise was created in Audition and saved in a raw PCM file.)

fid = fopen('WhiteNoise.pcm', 'r');

y = fread(fid, 'int16');

y = y(1:44100);

first = y(442:44100);

second = y(1:441);

y2 = cat(1, first, second);

y3 = y + y2;



You can see in Figure 7.42 that with a delay of 0.01 seconds, the frequencies that are combed out are 50 Hz, 150 Hz, 250 Hz, and so forth, while frequencies of 100 Hz, 200 Hz, 300 Hz, and so forth are boosted.

Figure 7.42 Comb filtering

Figure 7.42 Comb filtering

We created the comb filter above “by hand” by shifting the sample values and summing them. We can also use MATLAB’s fvtool.

a = [1  zeroes(1,440) 0.5];

b = [1];


axis([0 .008 -8 6]);


Figure 7.43 Comb filtering through fvtool

Figure 7.43 Comb filtering through fvtool

The x-axis in Figure 7.43 is normalized to the Nyquist frequency. That is, for a sampling rate of 44.1 kHz, the Nyquist frequency of 22,050 Hz is normalized to 1. Thus, 50 Hz falls at position 0.002 and 100 Hz at about 0.0045, etc.

Comb filters can be expressed as either non-recursive FIR filters or recursive IIR filters. A simple non-recursive comb filter is expressed by the following equation:

\begin{matrix}\mathbf{y}\left ( n \right )=\mathbf{x}\left ( n \right )+g \ast \mathbf{x}\left ( n-m \right )\\\textit{where }\mathbf{y}\textit{ is the filtered audio in the time domain}\\\mathbf{x}\textit{ is the original audio in the time domain}\\m\textit{ is the amount of delay in samples, and}\left | g \right |\leq 1\end{matrix}

Equation 7.16 Equation for an FIR comb filter

In an FIR comb filter, a fraction of an earlier sample is added to a later one, creating a repetition of the original sound. The distance between the two copies of the sound is controlled by m.

A simple recursive comb filter is expressed by the following equation:

\begin{matrix}\mathbf{y}\left ( n \right )=\mathbf{x}\left ( n \right )+g \ast \mathbf{y}\left ( n-m \right )\\\textit{where }\mathbf{y}\textit{ is the filtered audio in the time domain}\\\mathbf{x}\textit{ is the original audio in the time domain}\\m\textit{ is the amount of delay in samples, and}\left | g \right |\leq 1\end{matrix}

Equation 7.17 Equation for an IIR comb filter

You can see that Equation 7.17 represents a recursive filter because the output \mathbf{y}\left ( n-m \right ) is fed back in to create later output. With this feedback term, an earlier sound continues to have an effect on later sounds, but with decreasing intensity each time it is fed back in (because of the multiplier \left | g \right |\leq 1).

7.3.8 Flanging

Flanging is defined as a time-varying comb filter effect. It results from the application of a delay between two identical copies of audio where the amount of delay varies over time. This time-varying delay results in different frequencies and their harmonics being combed out a different moments in time. The time delay can be controlled by an LFO.

Flanging originated in the analog world as a trick performed with two tape machines playing identical copies of audio. The sound engineer was able to delay one of the two copies of the recording by pressing his finger on the rim or “flange” of the tape deck. Varying the pressure would vary the amount of delay, creating a “swooshing” sound.

Flanging effects have many variations. In basic flanging, the frequencies combed out are in a harmonic series at any moment in time. Flanging by means of phase-shifting can result in the combing of non-harmonic frequencies, such that the space between combed frequencies is not regular. The amount of feedback in flanging can also be varied. Flanging with a relatively long delay between copies of the audio is sometimes referred to a chorusing.

Guitars are a favorite instrument for flanging. The rock song “Barracuda” by Heart is good example of guitar flanging, but there are many others throughout rock history.

We leave flanging as an exercise for the reader, referring you to The Audio Programming Book cited in the references for an example implementation.

7.3.9 The Digital Signal Processing Toolkit in MATLAB

Section 7.3.2 gives you algorithms for creating a variety of FIR filters. MATLAB also provides built-in functions for creating FIR and IIR filters. Let's look at the IIR filters first.

MATLAB's butter function creates an IIR filter called a Butterworth filter, named for its creator. Consider this sequence of commands:

N = 10;

f = 0.5;

[a,b] = butter(N, f);


The butter function call sends in two arguments: the order of the desired filter, N; and the and the cutoff frequency, f. The cutoff frequency is normalized so that the Nyquist frequency (½ the sampling rate) is 1, and all valid frequencies lie between 0 and 1. The function call returns two vectors, a and b, corresponding to the vectors a and b in Equation 7.4. (For a simple low-pass filter, an order of 6 is fine.  The order is the number of coeffients.)

Now with the filter in hand, you can apply it using MATLAB’s filter function. The filter function takes the coefficients and the vector of audio samples as arguments:

output = filter(a,b,audio);


You can analyze the filter with the fvtool.



The fvtool provides a GUI through which you can see multiple views of the filter, including those in Figure 7.44. In the first figure, the blue line is the magnitude frequency response and the green line is phase.

Figure 7.44 Filter Visualization tool in MATLAB

Figure 7.44 Filter Visualization tool in MATLAB

Another way to create and apply an IIR filter in MATLAB is by means of the function yulewalk. Let's try a low-pass filter as a simple example. Figure 7.45 shows the idealized frequency response of a low-pass filter. The x-axis represents normalized frequencies, and f_c is the cutoff frequency. This particular filter allows frequencies that are up to ¼ the sampling rate to pass through, but filters out all the rest.

Figure 7.45 Frequency response of an ideal low-pass filter

Figure 7.45 Frequency response of an ideal low-pass filter

The first step in creating this filter is to store its "shape." This information is stored in a pair of parallel vectors, which we'll call f and m. For the four points on the graph in Figure 7.46, f stores the frequencies, and stores the corresponding magnitudes. That is, \mathbf{f}=\left [ f1\; f2\; f3\; f4 \right ] and \mathbf{m}=\left [ m1\; m2\; m3\; m4 \right ], as illustrated in the figure. For the example filter we have

f = [0 0.25 0.25 1];

m = [1 1 0 0];


Figure 7.46 Points corresponding to input parameters in yulewalk function

Figure 7.46 Points corresponding to input parameters in yulewalk function

Aside:  The yulewalk function in MATLAB is named for the Yule-Walker equations, a set of linear equations used in auto-regression modeling.

Now that you have an ideal response, you use the yulewalk function in MATLAB to determine what coefficients to use to approximate the ideal response.

[a,b] = yulewalk(N,f,m)


Again, an order N=6 filter is sufficient for the low-pass filter. You can use the same filter function as above to apply the filter. The resulting filter is given in Figure 7.47. Clearly, the filter cannot be perfectly created, as you can see from the large ripple after the cutoff point.

Figure 7.47 Frequency response of a yulewalk filter

Figure 7.47 Frequency response of a yulewalk filter

The finite counterpart to the yulewalk function is the fir2 function. Like butter, fir2 takes as input the order of the filter and two vectors corresponding to the shape of the filter's frequency response. Thus, we can use the same f and m as before. fir2 returns the vector h constituting the filter.

h = fir2(N,f,m);


We need to use a higher order filter because this is an FIR. N=30 is probably high enough.

MATLAB’s extensive signal processing toolkit includes a Filter Designer with an easy-to-use graphical user interface that allows you to design the types of filters discussed in this chapter. A Butterworth filter is shown in Figure 7.48. You can adjust the parameters in the tool and see how the roll-off and ripples change accordingly. Experimenting with this tool helps you know what to expect from filters designed by different algorithms.

Figure 7.48 Butterworth filter designed in MATLAB’s Filter Design and Analysis tool

Figure 7.48 Butterworth filter designed in MATLAB’s Filter Design and Analysis tool

7.3.10 Experiments with Filtering: Vocoders and Pitch Glides


Vocoders were introduced in Section 7.1.8. The implementation of a vocoder is sketched in Algorithm 7.6 and diagrammed in Figure 7.49. The MATLAB and C++ exercises associated with this section encourage you to try your hand at the implementation.

algorithm vocoder



c, an array of audio samples constituting the carrier signal

m, n array of audio samples constituting the modulator signal


v, the carrier wave modulated with the modulator wave */


Initialize v with 0s

Divide the carrier into octave-separated frequency bands with bandpass filters

Divide the modulator into the same octave-separated frequency bands with bandpass filters for each band

use the modulator as an amplitude envelope for the carrier


Algorithm 7.6 Sketch of an implementation of a vocoder

Figure 7.49 Overview of vocoder implementation

Figure 7.49 Overview of vocoder implementation

Another interesting programming exercise is implementation of a pitch glide. A Risset pitch glide is an audio illusion that sounds like a constantly rising pitch. It is the aural equivalent of the visual image of a stripe on a barber pole that seems to be rising constantly. Implementing the pitch glide is suggested as an exercise for this section.

7.3.11 Real-Time vs. Off-Line Processing

To this point, we’ve primarily considered off-line processing of audio data in the programs that we’ve asked you to write in the exercises.  This makes the concepts easier to grasp, but hides the very important issue of real-time processing, where operations have to keep pace with the rate at which sound is played.

Chapter 2 introduces the idea of audio streams.  In Chapter 2, we give a simple program that evaluates a sine function at the frequency of desire notes and writes the output directly to the audio device so that notes are played when the program runs.  Chapter 5 gives a program that reads a raw audio file and writes that to the audio device to play it as the program runs.  The program from Chapter 5 with a few modifications is given here for review.

/*Use option -lasound on compile line.  Send in number of samples and raw sound file name.*/

#include </usr/include/alsa/asoundlib.h>
#include <math.h>
#include <iostream>
using namespace std;

static char *device = "default";	/*default playback device */
snd_output_t *output = NULL;
#define PI 3.14159

int main(int argc, char *argv[])
        int err, numRead;
        snd_pcm_t *handle;
        snd_pcm_sframes_t frames;
        int numSamples = atoi(argv[1]);

        char* buffer = (char*) malloc((size_t) numSamples);
        FILE *inFile = fopen(argv[2], "rb");
        numRead = fread(buffer, 1, numSamples, inFile);

        if ((err = snd_pcm_open(&handle, device, SND_PCM_STREAM_PLAYBACK, 0)) < 0){
            printf("Playback open error: %s\n", snd_strerror(err));
        if ((err = snd_pcm_set_params(handle,
				44100, 1, 400000) ) < 0 ){
            printf("Playback open error: %s\n", snd_strerror(err));

        frames = snd_pcm_writei(handle, buffer, numSamples);
        if (frames < 0)
          frames = snd_pcm_recover(handle, frames, 0);
        if (frames < 0) {
          printf("snd_pcm_writei failed: %s\n", snd_strerror(err));

Program 7.1 Reading and writing raw audio data

This program uses the library function send_pcm_writei to send samples to the audio device to be played. The audio samples are read from in input file into a buffer and transmitted to the audio device without modification The variable buffer indicates where the samples are stored, and sizeof(buffer)/8 gives the number of samples given that this is 8-bit audio.

Consider what happens when you have a much larger stream of audio coming in and you want to process it in real time before writing it to the audio device. This entails continuously filling up and emptying the buffer at a rate that keeps up with the sampling rate.

Let’s do some analysis to determine how much time is available for processing based on a given buffer size. For a buffer size of N and a sampling rate of r, then N/r seconds can be passed before additional audio data will be required for playing. For and N=4096 and r=44100, this would be \frac{4096}{44100}=0.0929\: ms. (This scheme implies that there will be latency between the input and output, at most N/r seconds.)

What is you wanted to filter the input audio before sending it to the output? We’ve seen that filtering is more efficient in the frequency domain using the FFT. Assuming the input is in the time domain, our program has to do the following:

  • convert data to the frequency domain with inverse FFT
  • multiply the filter and the audio data
  • convert data back to the time domain with inverse FFT
  • write the data to the audio device

The computational complexity of the FFT and IFFT is 0\left ( N\log N \right ), on the order of 4096\ast 12=49152 operations (times 2). Multiplying the filter and the audio data is 0\left ( N \right ), and writing the data to the audio devices is also 0\left ( N \right ), adding on the order of the order of 2*4096 operations. This yields on the order of 106496 operations to be done in 0.0929\; ms, or about 0.9\; \mu s per operation. Considering that today’s computers can do more than 100,000 MIPS (millions of instructions per second), this is not unreasonable.

We refer the reader to Boulanger and Lazzarini's Audio Programming Book for more examples of real-time audio processing.

7.4 References

Flanagan, J. L., and R. M. Golden.  1966. "Phase Vocoder." Bell System Technical Journal. 45: 1493-1509.

Boulanger, Richard, and Victor Lazzarini, eds.  The Audio Programming Book. MIT Press, 2011.

Ifeachor, Emmanual C., and Barrie W. Jervis. Digital Signal Processsing: A Practical Approach. Addison-Wesley Publishing, 1993.