Sox in phonetic research

sox is a great command line tool for working with audio files, and this page illustrates some of the ways phoneticians can make use of it in their research. Once you have learned its syntax you can perform a wide variety of audio processing tasks that can easily be incorporated into a batch processing script to handle hundreds or thousands audio files.

In all of these examples, input and output files are assumed to be single-channel, unless specified otherwise. The Bash shell is also assumed, though most of the examples will work in other shells.

These are not the only possible usages, and you should consult the sox documentation for additional effects and options.

sox command line basics

The basic syntax of a sox command is:

   sox input1 [input2..n] output [effect]

An English translation of this is: 'Exchange this file [and these files] into that file [by means of some effect].' (The name sox derives from 'Sound eXchange'.) Portions of the command surrounded by [] are optional.

It's important to note that the sox command always looks for at least one input file and an output file, even when it's not obvious what those files should be. For example, when you want sox to generate one second of white noise and don't need to copy sample rate, etc. from some other file, then specifying an input file may not make sense. In such an event, you can substitute the special filename -n where the input file normally goes in the command line to tell sox not to look for a real input file.

Other times you might not want to create a real output file, for example, if you want the output of one sox command to be the input for a second sox command without creating an intermediate file. In that case you can use the special filename -p in place of the output filename to tell sox that the output will be piped to another command's input.

Converting audio file types, formats, and channels

Convert file type

This converts an .au file to .wav:

   sox input.au output.wav

Change sample rate

This resamples input.wav and creates a new file with a sample rate of 22050Hz:

   sox input.wav -r 22050 output.wav

Change sample rate and data size

This creates a new file with the specified sample rate and specifies that sample values will be 8 bits/sample.

   sox input.wav -r 22050 -b 8 output.wav

Change sample rate and file type:

This converts .wav to .aiff and resamples the signal to 12000Hz:

   sox input.wav -r 12000 output.aiff

Override a sample rate

Occasionally an audio file's header misidentifies the sample rate of the signal because of an error when the file was created or processed (or if the file is a raw audio file there is no header). You can override the header sample rate value by specifying the real sample rate as an input file option:

   sox -r 22050 input.wav output.wav

In this example the signal in output.wav contains the same number of samples as input.wav (i.e. it is *not* resampled), and the value of the sample rate property in the header will be 22050Hz, regardless of the sample rate value in input.wav's header.

Average two stereo channels into a single mono channel

   sox input.wav -c 1 output.wav avg

Here input.wav is a two-channel audio file. The -c 1 output format selects a single-channel output, and the avg effect averages the two input channels.

Remove a channel from a stereo recording to make a mono audio file

Some digital recorders always make stereo recordings, regardless of whether you use two microphones or not, resulting in a recording that either (1) has one channel with no audio other than amplifier noise; or (2) has two identical channels. In either case it is efficient to remove the unneeded channel from the recording--since your new file contains only half the data (not counting the file header), processing will be easier and faster, and you'll only need about half the storage to save the file.

To remove a channel you tell avg to copy only one of the channels:

   sox input.wav -c 1 output.wav avg -l

This copies only the left channel to the output file and drops the right one. Use avg -r to do the opposite.

Extracting a portion of an audio file

Use the trim effect to extract a portion of an audio file. For example, you can extract the first 10 seconds of an input file with:

   sox input.wav shorter.wav trim 0 10

The two numbers supplied to the trim effect define the start time and duration of the extracted audio. If the second number is prefixed with =, then the second number defines the end time of the extraction. The following commands are equivalent:

   sox input.wav shorter.wav trim 11.2 5        # extract 5 seconds of audio, starting at 11.2 seconds
   sox input.wav shorter.wav trim 11.2 =16.2    # extract audio from 11.2 to 16.2 seconds

If a time is prefixed with - it is interpreted relative to the end of the audio. This extracts the final 7.5 seconds:

   sox input.wav shorter.wav trim -7.5

As you can see, if no second number is provided sox extracts until the end of the file.

According to the sox documentation times are provided in hh:mm:ss.fraction format. In use we find that it is not necessary to convert seconds to minutes and hours, and these are equivalent:

   sox input.wav shorter.wav trim 139.2 88.3
   sox input.wav shorter.wav trim 02:19.2 01:28.3

Finally, you can extract using sample numbers instead of times by suffixing the number with s. These extract sample numbers 8000 to 10,000:

   sox input.wav shorter.wav trim 8000s 2000s
   sox input.wav shorter.wav trim 8000s =10000s

Synthesis

sox has a synth effect for generating tones or wideband noise that can be used in standalone form, or concatenated or mixed with an existing audio file.

Generate a periodic waveform

This generates a 2.25 second sine wave at 300Hz (sample rate 12050; 16-bit signed integer):

   sox -n -r 12050 -b 16 -s output.wav synth 2.25 sine 300

Generate the sum of two periodic waveforms

This chains two synth effects to add a a 250Hz square wave to the sine wave:

   sox -n -r 12050 -b 16 -s output.wav synth 2.25 sine 300 synth 2.25 square 250

See the sox documentation for additional options on creating swept rather than fixed frequency tones, and specifying phase, dc offset, and more.

Generate an aperiodic waveform (noise)

This produces a 3.5-second file with a 44.1KHz sample rate at 16 bits/sample. The '-n' parameter tells sox that no input file is used to determine the duration or data format:

   sox -n -r 44100 -b 16 noise.wav synth 3.5 whitenoise

You can also use noise with different spectral characteristics by using 'brownnoise' or 'pinknoise' instead of 'whitenoise'.

Generate an aperiodic waveform to match an audio file

You will often want to produce a noise file that matches the recording parameters of some other file, and for that you specify an input file:

   sox input.wav noise.wav synth whitenoise

This produces an output file that has the same number of samples (duration) and sample rate as the input file.

Add noise to an audio file

You can combine two sox commands to add noise to an input file in a new output file. These commands (1) generate noise matching the duration and sample rate of the input file and then (2) mix the generated noise with the input file to produce the new output:

   sox input.wav -p synth whitenoise vol 0.02 | sox -m input.wav - addednoise.wav

In the above example the '-p' parameter replaces the output file name in the first command and tells sox to pipe the noise output to the next sox command instead of to a file. This piped data stream shows up in the second command as the '-' argument, where the '-m' option tells sox to mix it with input.wav to produce the output file addednoise.wav.

It's important to note that the above example employs the simplest kind of mixing in which values of the two input files are simply added together, which introduces the risk of clipping in your output file. This clipping risk might not be a problem if you know that the amplitude range of your input file leaves enough room for the noise that you want to add. For instance, if the amplitude values in your input file have a maximum/minimum of +-0.6 you can add noise with a range of +-0.4 with no danger of clipping. If you want to add more noise than that you will need to apply the 'vol' gain effect or the 'norm' effect to prevent clipping.

An alternative syntax for the previous command is to use the shell's process substitution construct <() to create the noise as the second input file to sox -m:

   sox -m input.wav <(sox input.wav -p synth whitenoise vol 0.02) addednoise.wav

Adjusting the volume

The vol effect

The vol effect amplifies or attenuates the audio signal.

By default sox applies a linear scale to the signal. vol values greater than 1 result in amplification, and values between 0 and 1 result in attenuation. This example doubles the sample values of a signal:

   sox input.wav output.wav vol 2

One way to maximize the amplitude of a signal without clipping is to use the stat effect to retrieve a multiplier that will result in a maximized signal and pass it to the vol effect:

   sox input.wav output.wav vol $(sox input.wav -n stat -v 2>&1)

In the above example two calls to sox are necessary. The second call is performed via the shell's command substitution syntax $() and retrieves the multiplier (which sox outputs to STDERR, which we redirect to STDOUT with the 2>&1 syntax), and this multiplier is used as the literal value of the vol effect in the first command.

Note that the above complex command achieves the same result as the simple norm effect:

   sox input.wav output.wav norm

The complex command may be faster than norm since it does not create a temporary file during processing, which norm does. For long audio files this advantage could be considerable. The complex command might not work for some types of encodings (PCM should be fine), and the 2>&1 syntax is not available in all shells.

Scale the amplitude of an audio file so that the max/min values fit within a specific portion of the possible range

Here's a more complicated example. Say we want to scale our audio signal so that sample values are in the range +/-0.8 rather than the allowed maximum range +/-1.0, perhaps because we want to add noise with an amplitude range of +/-0.2. To do this we can use the stat effect to grab the vol multiplier and multiply it by 0.8:

   sox input.wav output.wav vol $(echo "$(sox input.wav -n stat -v 2>&1) * 0.8" | bc -l)

This example makes use of nested command substitution to first retrieve, then scale, the multiplier, which is subsequently used as the literal value of the vol effect. The bc calculator must be available on your system in order to perform the multiplication.

Add noise to an audio file without clipping

Putting things all together, here's one way to add noise to an audio file and be sure to avoid clipping. Note that the command is broken into multiple lines for readability. The \ character tells the shell that the command continues on the following line.

   sf1=0.94; sf2=0.05; \
     sox -m \
     -v $(echo "$(sox input.wav -n stat -v 2>&1) * ${sf1}" | bc -l)  input.wav \
     -v ${sf2} <(sox input.wav -p synth whitenoise) \
     -b 16 output.wav

Let's break this down. This complex command mixes two input files produced by subcommands. The basic form of a mixing command is:

   sox -m [-v factor1] input1.wav [-v factor2] input2.wav output.wav

Here input1.wav and input2.wav are mixed to produce output.wav. If provided, the -v factorN arguments linearly scale the corresponding input files.

In our complex command the first line

   sf1=0.94; sf2=0.05;

sets the value of two shell variables, the scaling factors sf1 and sf2. These are used to scale the amplitudes of the two input files.

The second line

   sox -m

tells sox to look for multiple input files to mix into an output file.

The third line identifies the first input file, which will be an amplitude-scaled version of our real input file

   -v $(echo "$(sox input.wav -n stat -v 2>&1) * ${sf1}" | bc -l)  input.wav

As we saw earlier, the $(echo "$(sox input.wav -n stat -v 2>&1) * ${sf1}" | bc -l) construct creates a gain multiplier that scales a signal to a percentage (${sf1}; 0.94) of the maximum possible amplitude. This multiplier is used as the literal value of the -v argument, scaling input.wav so that it uses 94% of the allowed amplitude range.

The next line creates and scales a noise input based on the duration and sample rate of input.wav:

   -v ${sf2} <(sox input.wav -p synth whitenoise)

Here the shell's process substitution construct <() is used to chain the output of a nested command to the input of the enclosing sox -m command. The special output filename -p is used in place of a real output filename so that the output data can be used as input to the enclosing sox -m command. Since synth generates full-amplitude noise by default, we use -v ${sf2} to scale the noise to 5% (0.05) of the maximum. (Note the subtle difference between process substitution, used in this line, and command substitution $(), used in the preceding line. Command substitution produces a string to be used as a literal value in the enclosing command. Process substitution is used to pipe data to the enclosing command. The point is subtle because sox can use either a literal filename or a data stream as an input file.)

The final line identifies the output filename and data size:

   -b 16 output.wav

The -b 16 format is important here, as your output will have a data size of 32 bits/sample without it, which is more than you need, and other software might not be able to deal with the deeper bit depth. The reason this happens is that the -p special filename pipes data using sox's internal 32-bit format, and this makes the mixing command default to the same.