将24位音频样本转换为int32数组（小尾数WAV文件）

Question

I read a 24-bit mono audio .wav file into an array of type <i4 ( <i3 doesn't exist)

data = numpy.fromfile(fid, dtype=`<i4`, count=size//3)

What should I do in order to get the audio samples properly ? Should I swap bytes order of something like this, how ?

Answer 1

You can convert the data into a numpy array of uint8 , then add the 0 to each sample by using reshape and hstack ;

In [1]: import numpy as np

I'm using a generated sequence here as an example.

In [2]: a = np.array([1,2,3]*10, dtype=np.uint8)

In [3]: a
Out[3]: 
array([1, 2, 3, 1, 2, 3, 1, 2, 3, 1, 2, 3, 1, 2, 3, 1, 2, 3, 1, 2, 3, 1, 2,
       3, 1, 2, 3, 1, 2, 3], dtype=uint8)

In [4]: a = a.reshape((-1,3))

Reshape allows you to group the samples:

In [5]: a
Out[5]: 
array([[1, 2, 3],
       [1, 2, 3],
       [1, 2, 3],
       [1, 2, 3],
       [1, 2, 3],
       [1, 2, 3],
       [1, 2, 3],
       [1, 2, 3],
       [1, 2, 3],
       [1, 2, 3]], dtype=uint8)

Make the zeros that have to be added.

In [6]: b = np.zeros(10, dtype=np.uint8).reshape((-1,1))

In [7]: b
Out[7]: 
array([[0],
       [0],
       [0],
       [0],
       [0],
       [0],
       [0],
       [0],
       [0],
       [0]], dtype=uint8)

Now we add the zeroes. Assuming you're using a little-endian system, the added zero goes at the front, to scale the data.

(I hope I got this endianness stuff right. If the sample now sounds very faint, I got it wrong and you need to use (a,b) instead of (b,a) )

In [8]: c = np.hstack((b, a))

In [9]: c
Out[9]: 
array([[0, 1, 2, 3],
       [0, 1, 2, 3],
       [0, 1, 2, 3],
       [0, 1, 2, 3],
       [0, 1, 2, 3],
       [0, 1, 2, 3],
       [0, 1, 2, 3],
       [0, 1, 2, 3],
       [0, 1, 2, 3],
       [0, 1, 2, 3]], dtype=uint8)

Reshape it back.

In [10]: c.reshape((1,-1))
Out[10]: 
array([[0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1,
        2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3]], dtype=uint8)

Convert to bytes:

In [11]: bytearray(c.reshape((1,-1)))
bytearray(b'\x00\x01\x02\x03\x00\x01\x02\x03\x00\x01\x02\x03\x00\x01\x02\x03\x00\x01\x02\x03\x00\x01\x02\x03\x00\x01\x02\x03\x00\x01\x02\x03\x00\x01\x02\x03\x00\x01\x02\x03')

Now you have 4-byte samples.

Answer 2

Here is the solution for reading 24 bits files (thanks to Warren Weckesser's gist https://gist.github.com/WarrenWeckesser/7461781 ) :

data = numpy.fromfile(fid, dtype='u1', count=size) # first read byte per byte

a = numpy.empty((len(data)/3, 4), dtype=`u1`)
a[:, :3] = data.reshape((-1, 3))
a[:, 3:] = (a[:, 3 - 1:3] >> 7) * 255
data = a.view('<i4').reshape(a.shape[:-1])

This can be directly inserted in def _read_data_chunk(fid, noc, bits): ( scipy\\io\\wavfile.py ).