I want to generate a 24-bit WAV-format audio file using Python 2.7 from an array of floating point values between -1 and 1. I can't use scipy.io.wavfile.write because it only supports 16 or 32 bits. The documentation for Python's own wave module doesn't specify what format of data it takes.
So is it possible to do this in Python?
Another option is available in wavio (also on PyPI: https://pypi.python.org/pypi/wavio), a small module I created as a work-around to the problem of scipy not yet supporting 24 bit WAV files. The file wavio.py contains the function write, which writes a numpy array to a WAV file. To write a 24-bit file, use the argument sampwidth=3. The only dependency of wavio is numpy; wavio uses the standard library wave to deal with the WAV file format.
For example,
In [21]: import numpy as np
In [22]: import wavio
In [23]: rate = 22050 # samples per second
In [24]: T = 3 # sample duration (seconds)
In [25]: f = 440.0 # sound frequency (Hz)
In [26]: t = np.linspace(0, T, T*rate, endpoint=False)
In [27]: sig = np.sin(2 * np.pi * f * t)
In [28]: wavio.write("sine24.wav", sig, rate, sampwidth=3)
I already submitted an answer to this question 2 years ago, where I recommended scikits.audiolab.
In the meantime, the situation has changed and now there is a library available which is much easier to use and much easier to install, it even comes with its own copy of the libsndfile library for Windows and OSX (on Linux it's easy to install anyway): PySoundFile!
If you have CFFI and NumPy installed, you can install PySoundFile simply by running
pip install soundfile --user
Writing a 24-bit WAV file is easy:
import soundfile as sf
sf.write('my_24bit_file.wav', my_audio_data, 44100, 'PCM_24')
In this example, my_audio_data has to be a NumPy array with dtype 'float64', 'float32', 'int32' or 'int16'.
BTW, I made an overview page where I tried to compare many available Python libraries for reading/writing sound files.
Try the wave module:
In [1]: import wave
In [2]: w = wave.open('foo.wav', 'w') # open for writing
In [3]: w.setsampwidth(3) # 3 bytes/sample
Python can only pack integers in 2 and 4 bite sizes. So you can use a numpy array with a dtype on int32, and use a list comprehension to get 3/4 of the bytes of each integer:
In [14]: d = np.array([1,2,3,4], dtype=np.int32)
In [15]: d
Out[15]: array([1, 2, 3, 4], dtype=int32)
In [16]: [d.data[i:i+3] for i in range(0,len(d)*d.dtype.itemsize, d.dtype.itemsize)]
Out[16]: ['\x01\x00\x00', '\x02\x00\x00', '\x03\x00\x00', '\x04\x00\x00']
Using the wave module, the Wave_write.writeframes function expects WAV data to be packed into a 3-byte string in little-endian format. The following code does the trick:
import wave
from contextlib import closing
import struct
def wavwrite_24(fname, fs, data):
data_as_bytes = (struct.pack('<i', int(samp*(2**23-1))) for samp in data)
with closing(wave.open(fname, 'wb')) as wavwriter:
wavwriter.setnchannels(1)
wavwriter.setsampwidth(3)
wavwriter.setframerate(fs)
for data_bytes in data_as_bytes:
wavwriter.writeframes(data_bytes[0:3])
You should try scikits.audiolab:
import numpy as np
from scikits.audiolab import Sndfile, Format
sig = np.array([0, 1, 0, -1, 0], dtype=np.float32)
f = Sndfile('test_pcm24.wav', 'w', Format('wav', 'pcm24'), 1, 44100)
f.write_frames(sig)
f.close() # use contextlib.closing in real code
And to read it again:
f = Sndfile('test_pcm24.wav')
sig = f.read_frames(f.nframes, dtype=np.float32)
f.close() # use contextlib.closing in real code
scikits.audiolab uses libsndfile, so in addition to WAV files, you can also use FLAC, OGG and some more file formats.
