Create 32bit float wav file in python?

Question

I want to create 32bit float WAV files in Python (2.x). While "standard" WAV files usually use int, many professional audio applications process (and save) audio data as float. The standard wave module is not capable to do this: http://bugs.python.org/issue16525 Has anyone achieved this without using patched modules ? tnx for any help.

Answer 1

Here's my contribution... includes arbitrary word size and arbitrary number of channels. I've taken the liberty of changing float32_wav_file to include a file save for testing. Note that the multichannel data portion of the file structure interleaved. That loop could be be greatly pythonized I'm sure.

# see http://stackoverflow.com/questions/15576798/create-32bit-float-wav-file-in-python
# see... http://blog.theroyweb.com/extracting-wav-file-header-information-using-a-python-script
import struct
def float32_wav_file(file_name, sample_array, sample_rate):
    (M,N)=sample_array.shape
    #print "len sample_array=(%d,%d)" % (M,N)
    byte_count = M * N * 4 # (len(sample_array)) * 4  # 32-bit floats
    wav_file = ""
    # write the header
    wav_file += struct.pack('<ccccIccccccccIHHIIHH',
        'R', 'I', 'F', 'F',
        byte_count + 0x2c - 8,  # header size
        'W', 'A', 'V', 'E', 'f', 'm', 't', ' ',
        0x10,  # size of 'fmt ' header
        3,  # format 3 = floating-point PCM
        M,  # channels
        sample_rate,  # samples / second
        sample_rate * 4,  # bytes / second
        4,  # block alignment
        32)  # bits / sample
    wav_file += struct.pack('<ccccI',
        'd', 'a', 't', 'a', byte_count)
    print "packing..."
    for j in range(0,N):
        for k in range(0,M):
            wav_file += struct.pack("<f", sample_array[k,j])
    print "saving..."
    fi=open(file_name,'wb')
    for value in wav_file:
        fi.write(value)
    fi.close()

    return wav_file

import numpy as np
def wav_file_read(filename):
    fi=open(filename,'rb')
    data=fi.read()
    fi.close()
    A, B, C,  D        =struct.unpack('4c', data[0:4])      # 'RIFF'
    ChunkSize     =struct.unpack('<l', data[4:8])[0]   #4+(8+SubChunk1Size)+8+SubChunk2Size)
    A, B, C,  D        =struct.unpack('4c', data[8:12])     # 'WAVE'
    A, B, C,  D        =struct.unpack('4c', data[12:16])    # 'fmt '
    Subchunk1Size =struct.unpack('<l', data[16:20])[0] # LITTLE ENDIAN, long, 16
    AudioFormat   =struct.unpack('<h', data[20:22])[0] # LITTLE ENDIAN, short, 1
    NumChannels   =struct.unpack('<h', data[22:24])[0] # LITTLE ENDIAN, short, Mono = 1, Stereo = 2
    SampleRate    =struct.unpack('<l', data[24:28])[0] # LITTLE ENDIAN, long,  sample rate in samples per second
    ByteRate      =struct.unpack('<l', data[28:32])[0] # self.SampleRate * self.NumChannels * self.BitsPerSample/8)) # (ByteRate) LITTLE ENDIAN, long
    BlockAlign    =struct.unpack('<h', data[32:34])[0] # self.NumChannels * self.BitsPerSample/8))  # (BlockAlign) LITTLE ENDIAN, short
    BitsPerSample =struct.unpack('<h', data[34:36])[0] # LITTLE ENDIAN, short
    A,  B,  C,  D      =struct.unpack('4c', data[36:40])    # BIG ENDIAN, char*4
    SubChunk2Size      =struct.unpack('<l', data[40:44])[0] # LITTLE ENDIAN, long
    waveData=data[44:]
    (M,N)=(len(waveData),len(waveData[0]))   

    print("ChunkSize     =%d\nSubchunk1Size =%d\nAudioFormat   =%d\nNumChannels   =%d\nSampleRate    =%d\nByteRate      =%d\nBlockAlign    =%d\nBitsPerSample =%d\nA:%c,  B:%c,  C:%c,  D:%c\nSubChunk2Size =%d" %
        (ChunkSize     , 
        Subchunk1Size, 
        AudioFormat   ,
        NumChannels   ,
        SampleRate    ,
        ByteRate      ,
        BlockAlign    ,
        BitsPerSample ,
        A,  B,  C,  D ,
        SubChunk2Size ))    


    if BitsPerSample==8:
        print "Unpacking 8 bits on len(waveData)=%d" % len(waveData)
        d=np.fromstring(waveData,np.uint8)
        floatdata=d.astype(np.float64)/np.float(127)    
    elif BitsPerSample==16:
        print "Unpacking 16 bits on len(waveData)=%d" % len(waveData)
        d=np.zeros(SubChunk2Size/2, dtype=np.int16)
        j=0
        for k in range(0, SubChunk2Size, 2):
            d[j]=struct.unpack('<h',waveData[k:k+2])[0]
            j=j+1
        floatdata=d.astype(np.float64)/np.float(32767)    
    elif BitsPerSample==24:
        print "Unpacking 24 bits on len(waveData)=%d" % len(waveData)
        d=np.zeros(SubChunk2Size/3,  dtype=np.int32)
        j=0
        for k in range(0, SubChunk2Size, 3):
            d[j]=struct.unpack('<l',struct.pack('c',waveData[k])+waveData[k:k+3])[0]
            j=j+1
        floatdata=d.astype(np.float64)/np.float(2147483647)    
    else: # anything else will be considered 32 bits
        print "Unpacking 32 bits on len(waveData)=%d" % len(waveData)
        d=np.fromstring(waveData,np.int32)
        floatdata=d.astype(np.float64)/np.float(2147483647)   
    v=floatdata[0::NumChannels]
    for i in range(1,NumChannels):
        v=np.vstack((v,floatdata[i::NumChannels]))
    #return (np.vstack((floatdata[0::2],floatdata[1::2])), SampleRate, NumChannels, BitsPerSample)
    return (v, SampleRate, NumChannels, BitsPerSample)

if __name__ == "__main__":
    (array,SampleRate,NumChannels,BitsPerSample)=wav_file_read("my_input_file.wav")
    wavefile=float32_wav_file("test_file.wav",array,SampleRate)

Answer 2

This sounded like fun (see my handle), so I hammered out something. Maybe you can use it. If your Python script generates a monophonic waveform of numerical values that fall between [-1.0 .. 1.0], send that waveform in through sample_array and also specify sample_rate (e.g., 44100 or 48000). This will return to you an array that you can write to disk as a .wav file.

I tested the resulting .wav output in Windows Media Player, Apple QuickTime Player, and VLC (all on Windows 7). They all played it.

def float32_wav_file(sample_array, sample_rate):
  byte_count = (len(sample_array)) * 4  # 32-bit floats
  wav_file = ""
  # write the header
  wav_file += struct.pack('<ccccIccccccccIHHIIHH',
    'R', 'I', 'F', 'F',
    byte_count + 0x2c - 8,  # header size
    'W', 'A', 'V', 'E', 'f', 'm', 't', ' ',
    0x10,  # size of 'fmt ' header
    3,  # format 3 = floating-point PCM
    1,  # channels
    sample_rate,  # samples / second
    sample_rate * 4,  # bytes / second
    4,  # block alignment
    32)  # bits / sample
  wav_file += struct.pack('<ccccI',
    'd', 'a', 't', 'a', byte_count)
  for sample in sample_array:
    wav_file += struct.pack("<f", sample)
  return wav_file