21.1. "audioop" — Manipulate raw audio data
*******************************************

The "audioop" module contains some useful operations on sound
fragments. It operates on sound fragments consisting of signed integer
samples 8, 16 or 32 bits wide, stored in Python strings.  This is the
same format as used by the "al" and "sunaudiodev" modules.  All scalar
items are integers, unless specified otherwise.

This module provides support for a-LAW, u-LAW and Intel/DVI ADPCM
encodings.

A few of the more complicated operations only take 16-bit samples,
otherwise the sample size (in bytes) is always a parameter of the
operation.

The module defines the following variables and functions:

exception audioop.error

   This exception is raised on all errors, such as unknown number of
   bytes per sample, etc.

audioop.add(fragment1, fragment2, width)

   Return a fragment which is the addition of the two samples passed
   as parameters. *width* is the sample width in bytes, either "1",
   "2" or "4".  Both fragments should have the same length.  Samples
   are truncated in case of overflow.

audioop.adpcm2lin(adpcmfragment, width, state)

   Decode an Intel/DVI ADPCM coded fragment to a linear fragment.  See
   the description of "lin2adpcm()" for details on ADPCM coding.
   Return a tuple "(sample, newstate)" where the sample has the width
   specified in *width*.

audioop.alaw2lin(fragment, width)

   Convert sound fragments in a-LAW encoding to linearly encoded sound
   fragments. a-LAW encoding always uses 8 bits samples, so *width*
   refers only to the sample width of the output fragment here.

   New in version 2.5.

audioop.avg(fragment, width)

   Return the average over all samples in the fragment.

audioop.avgpp(fragment, width)

   Return the average peak-peak value over all samples in the
   fragment. No filtering is done, so the usefulness of this routine
   is questionable.

audioop.bias(fragment, width, bias)

   Return a fragment that is the original fragment with a bias added
   to each sample.  Samples wrap around in case of overflow.

audioop.cross(fragment, width)

   Return the number of zero crossings in the fragment passed as an
   argument.

audioop.findfactor(fragment, reference)

   Return a factor *F* such that "rms(add(fragment, mul(reference,
   -F)))" is minimal, i.e., return the factor with which you should
   multiply *reference* to make it match as well as possible to
   *fragment*.  The fragments should both contain 2-byte samples.

   The time taken by this routine is proportional to "len(fragment)".

audioop.findfit(fragment, reference)

   Try to match *reference* as well as possible to a portion of
   *fragment* (which should be the longer fragment).  This is
   (conceptually) done by taking slices out of *fragment*, using
   "findfactor()" to compute the best match, and minimizing the
   result.  The fragments should both contain 2-byte samples. Return a
   tuple "(offset, factor)" where *offset* is the (integer) offset
   into *fragment* where the optimal match started and *factor* is the
   (floating-point) factor as per "findfactor()".

audioop.findmax(fragment, length)

   Search *fragment* for a slice of length *length* samples (not
   bytes!) with maximum energy, i.e., return *i* for which
   "rms(fragment[i*2:(i+length)*2])" is maximal.  The fragments should
   both contain 2-byte samples.

   The routine takes time proportional to "len(fragment)".

audioop.getsample(fragment, width, index)

   Return the value of sample *index* from the fragment.

audioop.lin2adpcm(fragment, width, state)

   Convert samples to 4 bit Intel/DVI ADPCM encoding.  ADPCM coding is
   an adaptive coding scheme, whereby each 4 bit number is the
   difference between one sample and the next, divided by a (varying)
   step.  The Intel/DVI ADPCM algorithm has been selected for use by
   the IMA, so it may well become a standard.

   *state* is a tuple containing the state of the coder.  The coder
   returns a tuple "(adpcmfrag, newstate)", and the *newstate* should
   be passed to the next call of "lin2adpcm()".  In the initial call,
   "None" can be passed as the state. *adpcmfrag* is the ADPCM coded
   fragment packed 2 4-bit values per byte.

audioop.lin2alaw(fragment, width)

   Convert samples in the audio fragment to a-LAW encoding and return
   this as a Python string.  a-LAW is an audio encoding format whereby
   you get a dynamic range of about 13 bits using only 8 bit samples.
   It is used by the Sun audio hardware, among others.

   New in version 2.5.

audioop.lin2lin(fragment, width, newwidth)

   Convert samples between 1-, 2- and 4-byte formats.

   Note:

     In some audio formats, such as .WAV files, 16 and 32 bit samples
     are signed, but 8 bit samples are unsigned.  So when converting
     to 8 bit wide samples for these formats, you need to also add 128
     to the result:

        new_frames = audioop.lin2lin(frames, old_width, 1)
        new_frames = audioop.bias(new_frames, 1, 128)

     The same, in reverse, has to be applied when converting from 8 to
     16 or 32 bit width samples.

audioop.lin2ulaw(fragment, width)

   Convert samples in the audio fragment to u-LAW encoding and return
   this as a Python string.  u-LAW is an audio encoding format whereby
   you get a dynamic range of about 14 bits using only 8 bit samples.
   It is used by the Sun audio hardware, among others.

audioop.max(fragment, width)

   Return the maximum of the *absolute value* of all samples in a
   fragment.

audioop.maxpp(fragment, width)

   Return the maximum peak-peak value in the sound fragment.

audioop.minmax(fragment, width)

   Return a tuple consisting of the minimum and maximum values of all
   samples in the sound fragment.

audioop.mul(fragment, width, factor)

   Return a fragment that has all samples in the original fragment
   multiplied by the floating-point value *factor*.  Samples are
   truncated in case of overflow.

audioop.ratecv(fragment, width, nchannels, inrate, outrate, state[, weightA[, weightB]])

   Convert the frame rate of the input fragment.

   *state* is a tuple containing the state of the converter.  The
   converter returns a tuple "(newfragment, newstate)", and *newstate*
   should be passed to the next call of "ratecv()".  The initial call
   should pass "None" as the state.

   The *weightA* and *weightB* arguments are parameters for a simple
   digital filter and default to "1" and "0" respectively.

audioop.reverse(fragment, width)

   Reverse the samples in a fragment and returns the modified
   fragment.

audioop.rms(fragment, width)

   Return the root-mean-square of the fragment, i.e.
   "sqrt(sum(S_i^2)/n)".

   This is a measure of the power in an audio signal.

audioop.tomono(fragment, width, lfactor, rfactor)

   Convert a stereo fragment to a mono fragment.  The left channel is
   multiplied by *lfactor* and the right channel by *rfactor* before
   adding the two channels to give a mono signal.

audioop.tostereo(fragment, width, lfactor, rfactor)

   Generate a stereo fragment from a mono fragment.  Each pair of
   samples in the stereo fragment are computed from the mono sample,
   whereby left channel samples are multiplied by *lfactor* and right
   channel samples by *rfactor*.

audioop.ulaw2lin(fragment, width)

   Convert sound fragments in u-LAW encoding to linearly encoded sound
   fragments. u-LAW encoding always uses 8 bits samples, so *width*
   refers only to the sample width of the output fragment here.

Note that operations such as "mul()" or "max()" make no distinction
between mono and stereo fragments, i.e. all samples are treated equal.
If this is a problem the stereo fragment should be split into two mono
fragments first and recombined later.  Here is an example of how to do
that:

   def mul_stereo(sample, width, lfactor, rfactor):
       lsample = audioop.tomono(sample, width, 1, 0)
       rsample = audioop.tomono(sample, width, 0, 1)
       lsample = audioop.mul(lsample, width, lfactor)
       rsample = audioop.mul(rsample, width, rfactor)
       lsample = audioop.tostereo(lsample, width, 1, 0)
       rsample = audioop.tostereo(rsample, width, 0, 1)
       return audioop.add(lsample, rsample, width)

If you use the ADPCM coder to build network packets and you want your
protocol to be stateless (i.e. to be able to tolerate packet loss) you
should not only transmit the data but also the state.  Note that you
should send the *initial* state (the one you passed to "lin2adpcm()")
along to the decoder, not the final state (as returned by the coder).
If you want to use "struct.Struct" to store the state in binary you
can code the first element (the predicted value) in 16 bits and the
second (the delta index) in 8.

The ADPCM coders have never been tried against other ADPCM coders,
only against themselves.  It could well be that I misinterpreted the
standards in which case they will not be interoperable with the
respective standards.

The "find*()" routines might look a bit funny at first sight. They are
primarily meant to do echo cancellation.  A reasonably fast way to do
this is to pick the most energetic piece of the output sample, locate
that in the input sample and subtract the whole output sample from the
input sample:

   def echocancel(outputdata, inputdata):
       pos = audioop.findmax(outputdata, 800)    # one tenth second
       out_test = outputdata[pos*2:]
       in_test = inputdata[pos*2:]
       ipos, factor = audioop.findfit(in_test, out_test)
       # Optional (for better cancellation):
       # factor = audioop.findfactor(in_test[ipos*2:ipos*2+len(out_test)],
       #              out_test)
       prefill = '\0'*(pos+ipos)*2
       postfill = '\0'*(len(inputdata)-len(prefill)-len(outputdata))
       outputdata = prefill + audioop.mul(outputdata, 2, -factor) + postfill
       return audioop.add(inputdata, outputdata, 2)
