soundfile 0.14.0

Breaking Changes

The soundfile module has evolved rapidly in the past. Most notably, we changed the import name from import pysoundfile to import soundfile in 0.7. In 0.6, we cleaned up many small inconsistencies, particularly in the the ordering and naming of function arguments and the removal of the indexing interface.

In 0.8.0, we changed the default value of always_2d from True to False. Also, the order of arguments of the write function changed from write(data, file, ...) to write(file, data, ...).

In 0.9.0, we changed the ctype arguments of the buffer_* methods to dtype, using the Numpy dtype notation. The old ctype arguments still work, but are now officially deprecated.

In 0.12.0, we changed the load order of the libsndfile library. Now, the packaged libsndfile in the platform-specific wheels is tried before falling back to any system-provided libsndfile. If you would prefer using the system-provided libsndfile, install the source package or source wheel instead of the platform-specific wheels.

Installation

The soundfile module depends on the Python packages CFFI and NumPy, and the library libsndfile.

In a modern Python, you can use pip install soundfile to download and install the latest release of the soundfile module and its dependencies. On Windows (64/32) and OS X (Intel/ARM) and Linux 64, this will also install a current version of the library libsndfile. If you install the source module, you need to install libsndfile using your distribution’s package manager, for example sudo apt install libsndfile1.

If you are running on an unusual platform or if you are using an older version of Python, you might need to install NumPy and CFFI separately, for example using the Anaconda package manager.

Building

Soundfile itself does not contain any compiled code and can be bundled into a wheel with the usual python setup.py bdist_wheel. However, soundfile relies on libsndfile, and optionally ships its own copy of libsndfile in the wheel.

To build a binary wheel that contains libsndfile, make sure to checkout and update the _soundfile_data submodule, then run python setup.py bdist_wheel as usual. If the resulting file size of the wheel is around one megabyte, a matching libsndfile has been bundled (without libsndfile, it’s around 25 KB).

To build binary wheels for all supported platforms, run python build_wheels.py, which will python setup.py bdist_wheel for each of the platforms we have precompiled libsndfiles for.

Error Reporting

In case of API usage errors the soundfile module raises the usual ValueError or TypeError.

For other errors SoundFileError is raised (used to be RuntimeError). Particularly, a LibsndfileError subclass of this exception is raised on errors reported by the libsndfile library. In that case the exception object provides the libsndfile internal error code in the LibsndfileError.code attribute and the raw libsndfile error message in the LibsndfileError.error_string attribute.

Read/Write Functions

Data can be written to the file using soundfile.write(), or read from the file using soundfile.read(). The soundfile module can open all file formats that libsndfile supports, for example WAV, FLAC, OGG and MAT files (see Known Issues below about writing OGG files).

Here is an example for a program that reads a wave file and copies it into an FLAC file:

import soundfile as sf

data, samplerate = sf.read('existing_file.wav')
sf.write('new_file.flac', data, samplerate)

Block Processing

Sound files can also be read in short, optionally overlapping blocks with soundfile.blocks(). For example, this calculates the signal level for each block of a long file:

import numpy as np
import soundfile as sf

rms = [np.sqrt(np.mean(block**2)) for block in
       sf.blocks('myfile.wav', blocksize=1024, overlap=512)]

SoundFile Objects

Sound files can also be opened as SoundFile objects. Every SoundFile has a specific sample rate, data format and a set number of channels.

If a file is opened, it is kept open for as long as the SoundFile object exists. The file closes when the object is garbage collected, but you should use the SoundFile.close() method or the context manager to close the file explicitly:

import soundfile as sf

with sf.SoundFile('myfile.wav', 'r+') as f:
    while f.tell() < f.frames:
        pos = f.tell()
        data = f.read(1024)
        f.seek(pos)
        f.write(data*2)

All data access uses frames as index. A frame is one discrete time-step in the sound file. Every frame contains as many samples as there are channels in the file.

RAW Files

soundfile.read() can usually auto-detect the file type of sound files. This is not possible for RAW files, though:

import soundfile as sf

data, samplerate = sf.read('myfile.raw', channels=1, samplerate=44100,
                           subtype='FLOAT')

Note that on x86, this defaults to endian='LITTLE'. If you are reading big endian data (mostly old PowerPC/6800-based files), you have to set endian='BIG' accordingly.

You can write RAW files in a similar way, but be advised that in most cases, a more expressive format is better and should be used instead.

Virtual IO

If you have an open file-like object, soundfile.read() can open it just like regular files:

import soundfile as sf
with open('filename.flac', 'rb') as f:
    data, samplerate = sf.read(f)

Here is an example using an HTTP request:

import io
import soundfile as sf
from urllib.request import urlopen

url = "http://tinyurl.com/shepard-risset"
data, samplerate = sf.read(io.BytesIO(urlopen(url).read()))

Note that the above example only works with Python 3.x. For Python 2.x support, replace the third line with:

from urllib2 import urlopen

import io
import soundfile as sf

def ogg2wav(ogg: bytes):
    ogg_buf = io.BytesIO(ogg)
    ogg_buf.name = 'file.ogg'
    data, samplerate = sf.read(ogg_buf)
    wav_buf = io.BytesIO()
    wav_buf.name = 'file.wav'
    sf.write(wav_buf, data, samplerate)
    wav_buf.seek(0)  # Necessary for `.read()` to return all bytes
    return wav_buf.read()

import soundfile as sf

# for example, this uncompressed 5 minute wav file with 32 kHz sample rate is 18 Mb
data, samplerate = sf.read('5min_32kHz.wav')

# maximum mp3 compression results in 1.1 Mb file, with either CONSTANT or VARIABLE bit rate
sf.write('max_compression_vbr.mp3', data, samplerate, bitrate_mode='VARIABLE', compression_level=.99)
sf.write('max_compression_cbr.mp3', data, samplerate, bitrate_mode='CONSTANT', compression_level=.99)

# minimum mp3 compression results in 3.5 Mb file
sf.write('min_compression_vbr.mp3', data, samplerate, bitrate_mode='VARIABLE', compression_level=0)

Thread Safety

The soundfile module is a thin wrapper around the low-level libsndfile C library. As such, soundfile provides the same thread safety guarantees as the underlying C library. See this libsndfile issue for more details.

In short, multithreaded use is safe as long as handles for readers or writers are not shared between threads. It is safe to concurrently open and read the same file through unique per-thread handles. It is not safe to concurrently write to the same file or share reader or writer handles. You may see garbage results or crashes if you do this.

See the advice in the Free-threaded Python guide on thread-safe Python programming for suggestions on how to synchronize access to a thread-unsafe object.

Known Issues

Writing to OGG files can result in empty files with certain versions of libsndfile. See #130 for news on this issue.

If using a Buildroot style system, Python has trouble locating libsndfile.so file, which causes python-soundfile to not be loaded. This is apparently a bug in python. For the time being, in soundfile.py, you can remove the call to _find_library and hardcode the location of the libsndfile.so in _ffi.dlopen. See #258 for discussion on this issue.

News

2013-08-27 V0.1.0 Bastian Bechtold:

Initial prototype. A simple wrapper for libsndfile in Python

2013-08-30 V0.2.0 Bastian Bechtold:

Bugfixes and more consistency with PySoundCard

2013-08-30 V0.2.1 Bastian Bechtold:

Bugfixes

2013-09-27 V0.3.0 Bastian Bechtold:

Added binary installer for Windows, and context manager

2013-11-06 V0.3.1 Bastian Bechtold:

Switched from distutils to setuptools for easier installation

2013-11-29 V0.4.0 Bastian Bechtold:

Thanks to David Blewett, now with Virtual IO!

2013-12-08 V0.4.1 Bastian Bechtold:

Thanks to Xidorn Quan, FLAC files are not float32 any more.

2014-02-26 V0.5.0 Bastian Bechtold:

Thanks to Matthias Geier, improved seeking and a flush() method.

2015-01-19 V0.6.0 Bastian Bechtold:

A big, big thank you to Matthias Geier, who did most of the work!

• Switched to float64 as default data type.

• Function arguments changed for consistency.

• Added unit tests.

• Added global read(), write(), blocks() convenience functions.

• Documentation overhaul and hosting on readthedocs.

• Added 'x' open mode.

• Added tell() method.

• Added __repr__() method.

2015-04-12 V0.7.0 Bastian Bechtold:

Again, thanks to Matthias Geier for all of his hard work, but also Nils Werner and Whistler7 for their many suggestions and help.

• Renamed import pysoundfile to import soundfile.

• Installation through pip wheels that contain the necessary libraries for OS X and Windows.

• Removed exclusive_creation argument to write().

• Added truncate() method.

2015-10-20 V0.8.0 Bastian Bechtold:

Again, Matthias Geier contributed a whole lot of hard work to this release.

• Changed the default value of always_2d from True to False.

• Numpy is now optional, and only loaded for read and write.

• Added SoundFile.buffer_read() and SoundFile.buffer_read_into() and SoundFile.buffer_write(), which read/write raw data without involving Numpy.

• Added info() function that returns metadata of a sound file.

• Changed the argument order of the write() function from write(data, file, ...) to write(file, data, ...)

And many more minor bug fixes.

2017-02-02 V0.9.0 Bastian Bechtold:

Thank you, Matthias Geier, Tomas Garcia, and Todd, for contributions for this release.

• Adds support for ALAC files.

• Adds new member __libsndfile_version__

• Adds number of frames to info class

• Adds dtype argument to buffer_* methods

• Deprecates ctype argument to buffer_* methods

• Adds official support for Python 3.6

And some minor bug fixes.

2017-11-12 V0.10.0 Bastian Bechtold:

Thank you, Matthias Geier, Toni Barth, Jon Peirce, Till Hoffmann, and Tomas Garcia, for contributions to this release.

• Should now work with cx_freeze.

• Several documentation fixes in the README.

• Removes deprecated ctype argument in favor of dtype in buffer_*().

• Adds SoundFile.frames in favor of now-deprecated __len__().

• Improves performance of blocks() and SoundFile.blocks().

• Improves import time by using CFFI’s out of line mode.

• Adds a build script for building distributions.

2022-06-02 V0.11.0 Bastian Bechtold:

Thank you, tennies, Hannes Helmholz, Christoph Boeddeker, Matt Vollrath, Matthias Geier, Jacek Konieczny, Boris Verkhovskiy, Jonas Haag, Eduardo Moguillansky, Panos Laganakos, Jarvy Jarvison, Domingo Ramirez, Tim Chagnon, Kyle Benesch, Fabian-Robert Stöter, Joe Todd

• MP3 support

• Adds binary wheels for macOS M1

• Improves compatibility with macOS, specifically for M1 machines

• Fixes file descriptor open for binary wheels on Windows and Python 3.5+

• Updates libsndfile to v1.1.0

• Adds get_strings method for retrieving all metadata at once

• Improves documentation, error messages and tests

• Displays length of very short files in samples

• Supports the file system path protocol (pathlib et al)

2023-02-02 V0.12.0 Bastian Bechtold

Thank you, Barabazs, Andrew Murray, Jon Peirce, for contributions to this release.

• Updated libsndfile to v1.2.0

• Improves precompiled library location, especially with py2app or cx-freeze.

• Now provide binary wheels for Linux x86_64

• Now prefers packaged libsndfile over system-installed libsndfile

2023-02-15 V0.12.1 Bastian Bechtold

Thank you, funnypig, for the bug report

• Fixed typo on library location detection if no packaged lib and no system lib was found

2025-01-02 V0.13.0 Bastian Bechtold

Thank you, Zhong Jianxin, mcclure, jneuendorf-i4h, aoirint, endolith, Guy Illes, ytya, Sam Lapp, Benjamin Moody

• Updated libsndfile to v1.2.2

• Linux arm64 builds added

• Numpy is now a dependency

• Fixed error in blocks, if file is very short

• Compression level and bitrate controls added for compressed files

• Various README improvements

• Various build system improvements

• Various improvements to error messages

2025-01-25 V0.13.1 Bastian Bechtold

Thank you, Brian McFee and Guy Illes

• Fixed regression in blocks

2026-06-06 V0.14.0 Bastian Bechtold

Thank you GesonAnko, Trevor Gamblin, Andreas Karatzas, Harish RS, Hunter Hogan

• Added type annotations

• Added Licensing note to wheel

• Fixed race condition when opening files concurrently

• Fixed regressions in test suite

• Removed support for Python <= 3.9

• Added ARM64 support for Windows