Running pesq on batches of audio

[giamic]

Hi all,
I'm working with batches of audio files stored in 2D numpy arrays with shape (batch_size, num_samples). Currently, the only way I found to run pesq through the batch is with a for loop but it is quite slow. Would it be possible to add a vectorized version of pesq that takes batches of audio in input and returns an array of pesq scores in output?

[ludlows]

thanks for your idea.
it is totally doable.
however, what behavior is supposed to be introduced once an error occurs in the batch processing?

[ludlows]

as you described, the proper solution is to utlize the mutiprocessing feature in python.

[ludlows]

hi @giamic ,
the multiprocessing feature is available now on dev branch.
you could try it with thousands of files if you are interested.
thanks.

[giamic]

This is great! On my local machine it took 4 times less to compute using the new algorithm.
I tested with the following code:

if __name__ == "__main__":
    fs = 16000
    duration = 10
    batch_size = 64
    input_shape = (batch_size, fs * duration)
    x = np.random.random(input_shape)
    y = np.random.random(input_shape)
    n_processors = multiprocessing.cpu_count()
    print(f"Available processors: {n_processors}")
    t0 = time.time()
    res = []
    for i in range(batch_size):
        res.append(pesq(fs, x[i], y[i], "wb"))
    print(f"Time without multithreading: {time.time() - t0:.3f} s")
    t1 = time.time()
    res_batch = pesq_batch(fs, x, y, "wb", n_processor=n_processors)
    print(f"Time with multithreading   : {time.time() - t1:.3f} s")
    assert np.allclose(res_batch, res)

and the output was:

Available processors: 8
Time without multithreading: 15.320 s
Time with multithreading   : 4.344 s

Process finished with exit code 0

This is brilliant. Thank you!

[giamic]

The only suggestion I would make is to use multiprocessing.cpu_count() as the default value for n_processor instead of None, but I guess that this is a matter of taste.

[ludlows]

thanks for your suggestion and the work on testing this feature.

about the default value of n_processor, I agree with your suggestion.