ref: 5bafb67083c07394616c04e67dab86a9ca40d698
dir: /README/
////////////////////////////////////////////////////////////////////////////
// **** AUDIO-STRETCH **** //
// Time Domain Harmonic Scaler //
// Copyright (c) 2022 David Bryant //
// All Rights Reserved. //
// Distributed under the BSD Software License (see license.txt) //
////////////////////////////////////////////////////////////////////////////
From Wikipedia, the free encyclopedia:
Time-domain harmonic scaling (TDHS) is a method for time-scale
modification of speech (or other audio signals), allowing the apparent
rate of speech articulation to be changed without affecting the
pitch-contour and the time-evolution of the formant structure. TDHS
differs from other time-scale modification algorithms in that
time-scaling operations are performed in the time domain (not the
frequency domain).
This project is an implementation of a TDHS library and a command-line demo
program to utilize it with standard WAV files. The command-line program
also incorporates silence detection so that can be handled differently.
There are two effects possible with TDHS and the audio-stretch demo. The
first is the more obvious mentioned above of changing the duration (or
speed) of a speech (or other audio) sample without modifying its pitch.
The other effect is similar, but after applying the duration change we
change the sampling rate in a complimentary manner to restore the original
duration and timing, which then results in the pitch being altered.
So when a ratio is supplied to the audio-stretch program, the default
operation is for the total duration of the audio file to be scaled by
exactly that ratio (0.5X to 2.0X), with the pitches remaining constant.
If the option to scale the sample-rate proportionally is specified (-s)
then the total duration and timing of the audio file will be preserved,
but the pitches will be scaled by the specified ratio instead. This is
useful for creating a "helium voice" effect and lots of other fun stuff.
Note that unless ratios of exactly 0.5 or 2.0 are used with the -s option,
non-standard sampling rates will probably result. Many programs will still
properly play these files, and audio editing programs will likely import
them correctly (by resampling), but it is possible that some applications
will barf on them. They can also be resampled to a standard rate using
an audio resampling tool I wrote that's also available here on GitHub:
https://github.com/dbry/audio-resampler
There's an option to cycle through the full possible ratio range in a
sinusoidal pattern, starting at 1.0, and either going up (-c) or down
(-cc) first. In this case any specified ratio is ignored (except if the
-s option is also specified to scale the sampling rate). The total period
is fixed at 2π seconds, at which point the output will again be exactly
aligned with the input.
*** Version 0.4 Enhancements ***
For version 0.4 two useful features were added. First, the ability to
cascade two instances of the stretcher was added. This is enabled by
including the flag STRETCH_DUAL_FLAG when initializing the stretcher
and allows double the stretch ratio of the regular code (i.e., now 0.25X
to 4.00X). Note that the audio quality degrades some when slowed beyond
2X, and generally voice becomes unintelligible when sped faster than 2X,
however these values may still be useful for some applications, and
specifically the very high speed values are useful for silence gaps
(see the next feature).
The other feature added is the ability to detect silence gaps in the
audio and apply a different (likely lower) stretch ratio to these areas.
This is currently not performed in the library itself, but in the demo
command-line program where it is highly configurable, but it should be
relatively easy to copy the functionality into another application. If
I get requests for it, I will consider moving it into the library.
There is a script to build the demo app on Linux (build.sh), and this also
allows building the app to test for UB (undefined behavior) and ASAN (bad
addressing). Also, some artificial test signals (both mono and stereo) and
a script (test.sh) for running them at various ratios has been added.
The current "help" display from the demo app:
AUDIO-STRETCH Time Domain Harmonic Scaling Demo Version 0.4
Copyright (c) 2022 David Bryant. All Rights Reserved.
Usage: AUDIO-STRETCH [-options] infile.wav outfile.wav
Options: -r<n.n> = stretch ratio (0.25 to 4.0, default = 1.0)
-g<n.n> = gap/silence stretch ratio (if different)
-u<n> = upper freq period limit (default = 333 Hz)
-l<n> = lower freq period limit (default = 55 Hz)
-b<n> = audio buffer/window length (ms, default = 25)
-t<n> = gap/silence threshold (dB re FS, default = -40)
-c = cycle through all ratios, starting higher
-cc = cycle through all ratios, starting lower
-d = force dual instance even for shallow ratios
-s = scale rate to preserve duration (not pitch)
-f = fast pitch detection (default >= 32 kHz)
-n = normal pitch detection (default < 32 kHz)
-q = quiet mode (display errors only)
-v = verbose (display lots of info)
-y = overwrite outfile if it exists
Web: Visit www.github.com/dbry/audio-stretch for latest version
Notes:
1. The program will handle only mono or stereo files in the WAV format. In
case of stereo, the two channels shouldn't be independent. The
audio must be 16-bit PCM and the acceptable sampling rates are from 8,000
to 48,000 Hz. Any additional RIFF info in the WAV file will be discarded.
The command-line program is only for little-endian architectures.
2. For stereo files, the pitch detection is done on a mono conversion of the
audio, but the scaling transformation is done on the independent channels.
If it is desired to have completely independent processing this can only
be done with two mono files. Note that this is not a limitation of the
library but of the demo utility (the library has no problem with multiple
contexts).
3. This technique (TDHS) is ideal for speech signals, but can also be used
for homophonic musical instruments. As the sound becomes increasingly
polyphonic, however, the quality and effectiveness will decrease. Also,
the period frequency limits provided by default are optimized for speech;
adjusting these may be required for best quality with non-speech audio.
4. The vast majority of the time required for TDHS is in the pitch detection,
and so this library implements two versions. The first is the standard
one that includes every sample and pitch period, and the second is an
optimized one that uses pairs of samples and only even pitch periods.
This second version is about 4X faster than the standard version, but
provides virtually the same quality. It is used by default for files with
sample rates of 32 kHz or higher, but its use can be forced on or off
from the command-line (see options above).