SoundTouch audio processing library v2.2

SoundTouch library Copyright © Olli Parviainen 2001-2020

1. Introduction

SoundTouch is an open-source audio processing library that allows changing the sound tempo, pitch and playback rate parameters independently from each other, i.e.:

1.1 Contact information

Author email: oparviai 'at' iki.fi

SoundTouch WWW page: http://soundtouch.surina.net

SoundTouch git repository: https://gitlab.com/soundtouch/soundtouch.git

2. Compiling SoundTouch

Before compiling, notice that you can choose the sample data format if it's desirable to use floating point sample data instead of 16bit integers. See section "sample data format" for more information.

Also notice that SoundTouch can use OpenMP instructions for parallel computation to accelerate the runtime processing speed in multi-core systems, however, these improvements need to be separately enabled before compiling. See OpenMP notes in Chapter 3 below.

2.1. Building in Microsoft Windows

Project files for Microsoft Visual C++ are supplied with the source code package. Go to Microsoft WWW page to download Microsoft Visual Studio Express version for free.

To build the binaries with Visual C++ compiler, either run "make-win.bat" script, or open the appropriate project files in source code directories with Visual Studio. The final executable will appear under the "SoundTouch\bin" directory. If using the Visual Studio IDE instead of the make-win.bat script, directories bin and lib may need to be created manually to the SoundTouch package root for the final executables. The make-win.bat script creates these directories automatically.

C# example: The source code package includes also a C# example application for Windows that shows how to invoke SoundTouch.dll dynamic-load library for processing mp3 audio.

OpenMP NOTE: If activating the OpenMP parallel computing in the compilation, the target program will require additional vcomp dll library to properly run. In Visual C++ 9.0 these libraries can be found in the following folders.

In Visual Studio 2008, a SP1 version may be required for these libraries. In other VC++ versions the required library will be expectedly found in similar "redist" location.

Notice that as minor demonstration of a "dll hell" phenomenon both the 32-bit and 64-bit version of vcomp90.dll have the same filename but different contents, thus choose the proper version to allow the program start.

2.2. Building in Gnu platforms

The SoundTouch library compiles in practically any platform supporting GNU compiler (GCC) tools. SoundTouch requires GCC version 4.3 or later.

To build and install the binaries, run the following commands in /soundtouch directory:

./bootstrap  -
 Creates "configure" file with local autoconf/automake toolset.
 
./configure  -

Configures the SoundTouch package for the local environment. Notice that "configure" file is not available before running the "./bootstrap" command as above.

 
make         -

Builds the SoundTouch library & SoundStretch utility. You can optionally add "-j" switch after "make" to speed up the compilation in multi-core systems.

 
make install -

Installs the SoundTouch & BPM libraries to /usr/local/lib and SoundStretch utility to /usr/local/bin. Please notice that 'root' privileges may be required to install the binaries to the destination locations.

2.2.1 Required GNU tools

Bash shell, GNU C++ compiler, libtool, autoconf and automake tools are required for compiling the SoundTouch library. These are usually included with the GNU/Linux distribution, but if not, install these packages first. For example, Ubuntu Linux can acquire and install these with the following command:

sudo apt-get install automake autoconf libtool build-essential

2.2.2 Problems with GCC compiler compatibility

At the release time the SoundTouch package has been tested to compile in GNU/Linux platform. However, If you have problems getting the SoundTouch library compiled, try disabling optimizations that are specific for x86 processors by running ./configure script with switch 

--enable-x86-optimizations=no
Alternatively, if you don't use GNU Configure system, edit file "include/STTypes.h" directly and remove the following definition:
#define SOUNDTOUCH_ALLOW_X86_OPTIMIZATIONS 1

2.2.3 Compiling portable Shared Library / DLL version in GNU environment

The GNU compilation does not automatically create a shared-library version of SoundTouch (.so or .dll) that features position-independent code and C-language api that are more suitable for cross-language development than C++ libraries.

Use script "make-gnu-dll-sh" to build a portable dynamic library version if such is desired.

2.3. Building in Android

Android compilation instructions are within the source code package, see file "source/Android-lib/README-SoundTouch-Android.html" in the source code package.

The Android compilation automatically builds separate .so library binaries for ARM, X86 and MIPS processor architectures. For optimal device support, include all these .so library binaries into the Android .apk application package, so the target Android device can automatically choose the proper library binary version to use.

The source/Android-lib folder includes also an Android example application that processes WAV audio files using SoundTouch library in Android devices.

2.4. Building in Mac

Install autoconf tool as instructed in http://macappstore.org/autoconf/.

Then, build as described above in section "Building in Gnu platforms".

3. About implementation & Usage tips

3.1. Supported sample data formats

The sample data format can be chosen between 16bit signed integer and 32bit floating point values.

The default sample type is 32bit floating point format, which also provides better sound quality than integer format because integer algorithms need to scale already intermediate calculation results to avoid integer overflows. These early integer scalings can slightly degrade output quality.

In Windows environment, the sample data format is chosen in file "STTypes.h" by choosing one of the following defines:

In GNU environment, the floating sample format is used by default, but integer sample format can be chosen by giving the following switch to the configure script: 

./configure --enable-integer-samples

The sample data can have either single (mono) or double (stereo) audio channel. Stereo data is interleaved so that every other data value is for left channel and every second for right channel. Notice that while it'd be possible in theory to process stereo sound as two separate mono channels, this isn't recommended because processing the channels separately would result in losing the phase coherency between the channels, which consequently would ruin the stereo effect.

Sample rates between 8000-48000H are supported.

3.2. Processing latency

The processing and latency constraints of the SoundTouch library are:

3.3. About algorithms

SoundTouch provides three seemingly independent effects: tempo, pitch and playback rate control. These three controls are implemented as combination of two primary effects, sample rate transposing and time-stretching.

Sample rate transposing affects both the audio stream duration and pitch. It's implemented simply by converting the original audio sample stream to the desired duration by interpolating from the original audio samples. In SoundTouch, linear interpolation with anti-alias filtering is used. Theoretically a higher-order interpolation provide better result than 1st order linear interpolation, but in audio application linear interpolation together with anti-alias filtering performs subjectively about as well as higher-order filtering would.

Time-stretching means changing the audio stream duration without affecting it's pitch. SoundTouch uses WSOLA-like time-stretching routines that operate in the time domain. Compared to sample rate transposing, time-stretching is a much heavier operation and also requires a longer processing "window" of sound samples used by the processing algorithm, thus increasing the algorithm input/output latency. Typical i/o latency for the SoundTouch time-stretch algorithm is around 100 ms.

Sample rate transposing and time-stretching are then used together to produce the tempo, pitch and rate controls:

3.4 Tuning the algorithm parameters

The time-stretch algorithm has few parameters that can be tuned to optimize sound quality for certain application. The current default parameters have been chosen by iterative if-then analysis (read: "trial and error") to obtain best subjective sound quality in pop/rock music processing, but in applications processing different kind of sound the default parameter set may result into a sub-optimal result.

The time-stretch algorithm default parameter values are set by the following #defines in file "TDStretch.h":

#define DEFAULT_SEQUENCE_MS     AUTOMATIC
#define DEFAULT_SEEKWINDOW_MS   AUTOMATIC
#define DEFAULT_OVERLAP_MS      8

These parameters affect to the time-stretch algorithm as follows:

Notice that these parameters can also be set during execution time with functions "TDStretch::setParameters()" and "SoundTouch::setSetting()".

The table below summaries how the parameters can be adjusted for different applications:

Parameter name Default value magnitude Larger value affects... Smaller value affects... Effect to CPU burden 
SEQUENCE_MS
 Default value is relatively large, chosen for slowing down music tempo Larger value is usually better for slowing down tempo. Growing the value decelerates the "echoing" artifact when slowing down the tempo. Smaller value might be better for speeding up tempo. Reducing the value accelerates the "echoing" artifact when slowing down the tempo Increasing the parameter value reduces computation burden 
SEEKWINDOW_MS
 Default value is relatively large, chosen for slowing down music tempo Larger value eases finding a good mixing position, but may cause a "drifting" artifact Smaller reduce possibility to find a good mixing position, but reduce the "drifting" artifact. Increasing the parameter value increases computation burden 
OVERLAP_MS
 Default value is relatively large, chosen to suit with above parameters. If you reduce the "sequence ms" setting, you might wish to try a smaller value. Increasing the parameter value increases computation burden

3.5 Performance Optimizations

Integer vs floating point:

Floating point sample type is generally recommended because it provides better sound quality.

However, execution speed difference between integer and floating point processing depends on the CPU architecture. As rule of thumb,

General optimizations:

The time-stretch routine has a 'quick' mode that substantially speeds up the algorithm but may slightly compromise the sound quality. This mode is activated by calling SoundTouch::setSetting() function with parameter id of SETTING_USE_QUICKSEEK and value "1", i.e.

setSetting(SETTING_USE_QUICKSEEK, 1);

CPU-specific optimizations:

Intel x86 specific SIMD optimizations are implemented using compiler intrinsics, providing about a 3x processing speedup for x86 compatible processors vs. non-SIMD implementation:

The algorithms are tuned to utilize autovectorization efficiently also in other CPU architectures, for example ARM cpus see approx 2.4x processing speedup when NEON SIMD support is present.

3.5 OpenMP parallel computation

SoundTouch 1.9 onwards support running the algorithms parallel in several CPU cores. Based on benchmark the experienced multi-core processing speed-up gain ranges between +30% (on a high-spec dual-core x86 Windows PC) to 215% (on a moderately low-spec quad-core ARM of Raspberry Pi2).

See an external blog article with more detailed discussion about the SoundTouch OpenMP optimization.

The parallel computing support is implemented using OpenMP spec 3.0 instructions. These instructions are supported by Visual C++ 2008 and later, and GCC v4.2 and later. Compilers that do not supporting OpenMP will ignore these optimizations and routines will still work properly. Possible warnings about unknown #pragmas are related to OpenMP support and can be safely ignored.

The OpenMP improvements are disabled by default, and need to be enabled by developer during compile-time. Reason for this is that parallel processing adds moderate runtime overhead in managing the multi-threading, so it may not be necessary nor desirable in all applications. For example real-time processing that is not constrained by CPU power will not benefit of speed-up provided by the parallel processing, in the contrary it may increase power consumption due to the increased overhead.

However, applications that run on low-spec multi-core CPUs and may otherwise have possibly constrained performance will benefit of the OpenMP improvements. This include for example multi-core embedded devices.

OpenMP parallel computation can be enabled before compiling SoundTouch library as follows:

4. SoundStretch audio processing utility

SoundStretch audio processing utility
Copyright (c) Olli Parviainen 2002-2015

SoundStretch is a simple command-line application that can change tempo, pitch and playback rates of WAV sound files. This program is intended primarily to demonstrate how the "SoundTouch" library can be used to process sound in your own program, but it can as well be used for processing sound files.

4.1. SoundStretch Usage Instructions

SoundStretch Usage syntax:

soundstretch infilename outfilename [switches]

Where: 

"infilename"
 Name of the input sound data file (in .WAV audio file format). Give "stdin" as filename to use standard input pipe. 
"outfilename"
 Name of the output sound file where the resulting sound is saved (in .WAV audio file format). This parameter may be omitted if you don't want to save the output (e.g. when only calculating BPM rate with '-bpm' switch). Give "stdout" as filename to use standard output pipe. 
[switches]
 Are one or more control switches.

Available control switches are:

-tempo=n
Change the sound tempo by n percents (n = -95.0 .. +5000.0 %) 
-pitch=n
 Change the sound pitch by n semitones (n = -60.0 .. + 60.0 semitones) 
-rate=n
 Change the sound playback rate by n percents (n = -95.0 .. +5000.0 %) 
-bpm=n
 Detect the Beats-Per-Minute (BPM) rate of the sound and adjust the tempo to meet 'n' BPMs. When this switch is applied, the "-tempo" switch is ignored. If "=n" is omitted, i.e. switch "-bpm" is used alone, then the BPM rate is estimated and displayed, but tempo not adjusted according to the BPM value. 
-quick
 Use quicker tempo change algorithm. Gains speed but loses sound quality. 
-naa
 Don't use anti-alias filtering in sample rate transposing. Gains speed but loses sound quality. 
-license
 Displays the program license text (LGPL)

Notes:

4.2. SoundStretch usage examples

Example 1

The following command increases tempo of the sound file "originalfile.wav" by 12.5% and stores result to file "destinationfile.wav":

soundstretch originalfile.wav destinationfile.wav -tempo=12.5

Example 2

The following command decreases the sound pitch (key) of the sound file "orig.wav" by two semitones and stores the result to file "dest.wav":

soundstretch orig.wav dest.wav -pitch=-2

Example 3

The following command processes the file "orig.wav" by decreasing the sound tempo by 25.3% and increasing the sound pitch (key) by 1.5 semitones. Resulting .wav audio data is directed to standard output pipe:

soundstretch orig.wav stdout -tempo=-25.3 -pitch=1.5

Example 4

The following command detects the BPM rate of the file "orig.wav" and adjusts the tempo to match 100 beats per minute. Result is stored to file "dest.wav":

soundstretch orig.wav dest.wav -bpm=100

Example 5

The following command reads .wav sound data from standard input pipe and estimates the BPM rate:

soundstretch stdin -bpm

Example 6

The following command tunes song from original 440Hz tuning to 432Hz tuning: this corresponds to lowering the pitch by -0.318 semitones:

soundstretch original.wav output.wav -pitch=-0.318

5. Change History

5.1. SoundTouch library Change History

2.2:

2.1.2:

2.1.1:

2.1:

2.0:

1.9.2:

1.9.1:

1.9:

1.8.0:

1.7.1:

1.7.0:

1.6.0:

1.5.0:

1.4.1:

1.4.0:

1.3.1:

1.3.0:

1.2.1:

1.2.0:

1.1.1:

1.0.1:

1.0:

5.2. SoundStretch application Change History

1.9:

1.7.0:

1.5.0:

1.4.0:

1.3.0:

1.2.1:

1.2.0:

1.1.1:

1.1:

1.01:

6. Acknowledgements

Kudos for these people who have contributed to development or submitted bugfixes:

Moral greetings to all other contributors and users also!

7. LICENSE

SoundTouch audio processing library
Copyright (c) Olli Parviainen

This library is free software; you can redistribute it and/or modify it under the terms of the GNU Lesser General Public License version 2.1 as published by the Free Software Foundation.

This library is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more details.

You should have received a copy of the GNU Lesser General Public License along with this library; if not, write to the Free Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA

---

commercial license alternative also available, contact author for details.