Unable to use Multithread for librosa melspectrogram

Question

I have over 1000 audio files (it's just a initial development, in the future, there will be even more audio files), and would like to convert them to melspectrogram.

Since my workstation has a Intel® Xeon® Processor E5-2698 v3, which has 32 threads, I would like to use multithread to do my job.

My code

import os
import librosa
from librosa.display import specshow
from natsort import natsorted
import numpy as np
import sys 
# Libraries for multi thread
from multiprocessing.dummy import Pool as ThreadPool
import subprocess
pool = ThreadPool(20) 

songlist = os.listdir('../opensmile/devset_2015/')
songlist= natsorted(songlist)

def get_spectrogram(song):
    print("start")
    y, sr = librosa.load('../opensmile/devset_2015/' + song)

    ## Add some function to cut y
    y_list = y
    ##

    for i, y_i in enumerate([y_list]): # can remove for loop if no audio is cut
        S = librosa.feature.melspectrogram(y=y, sr=sr, n_mels=128,fmax=8000)
        try:
            np.save('./Test/' + song + '/' + str(i), S)
        except:
            os.makedirs('./Test/' + song)
            np.save('./Test/' + song + '/' + str(i), S)
        print("done saving")

pool.map(get_spectrogram, songlist)

My Problem

However, my script freezes after finished the first conversion.

To debug what's going on, I commented out S = librosa.feature.melspectrogram(y=y, sr=sr, n_mels=128,fmax=8000) and replace it by S=0. Then the multi-thread code works fine.

What's wrong with the librosa.feature.melspectrogram function? Does it not support multi-thread? Or is it a problem of ffmpeg? (When using librosa, it asks me to install ffmpeg before.)

Answer 1

I recommend using joblib to parallel process with librosa. I believe librosa is using it internally, so this might avoid some conflicts. Below is a working example, based on code that I regularly use to process some 10k files.

import os.path
import joblib
import librosa
import numpy

def compute(inpath, outpath):
    y, sr = librosa.load(inpath)
    S = librosa.feature.melspectrogram(y=y, sr=sr, n_mels=128, fmax=8000)
    numpy.save(outpath, S)
    return outpath

out_dir = 'temp/'
n_jobs=8
verbose=1

# as an reproducable example just processes the same input file
# but making sure to give them unique output names
inputs = [ librosa.util.example_audio_file() ] * 10
outputs = [ os.path.join(out_dir, '{}.npy'.format(n)) for n in range(len(inputs)) ]

jobs = [ joblib.delayed(compute)(i, o) for i,o in zip(inputs, outputs) ]
out = joblib.Parallel(n_jobs=n_jobs, verbose=verbose)(jobs)

print(out)

Output

[Parallel(n_jobs=8)]: Using backend LokyBackend with 8 concurrent workers.
[Parallel(n_jobs=8)]: Done   6 out of  10 | elapsed:   10.4s remaining:    6.9s
[Parallel(n_jobs=8)]: Done  10 out of  10 | elapsed:   13.2s finished
['temp/0.npy', 'temp/1.npy', 'temp/2.npy', 'temp/3.npy', 'temp/4.npy', 'temp/5.npy', 'temp/6.npy', 'temp/7.npy', 'temp/8.npy', 'temp/9.npy']

Answer 2

Another solution is to use Tacotron2 implementation to extract melspectrogram, which is multithread compatible. https://github.com/NVIDIA/tacotron2/blob/master/data_utils.py