Merge pull request SeanNaren#496 from shamoons/495/soundfile

Using soundfile.read to include more than just .wav files

.gitignore

@@ -104,3 +104,4 @@ ENV/
 models/
 data/*.csv
 data/*_dataset/
+.vscode/

data/data_loader.py

@@ -5,9 +5,9 @@
 import librosa
 import numpy as np
 import scipy.signal
+import soundfile as sf
 import sox
 import torch
-from scipy.io.wavfile import read
 from torch.utils.data import Dataset, Sampler, DistributedSampler, DataLoader
 
 from .spec_augment import spec_augment
@@ -21,7 +21,9 @@
 
 
 def load_audio(path):
- sample_rate, sound = read(path)
+ sound, sample_rate = sf.read(path, dtype='int16')
+ # TODO this should be 32768.0 to get twos-complement range.
+ # TODO the difference is negligible but should be fixed for new models.
  sound = sound.astype('float32') / 32767 # normalize audio
  if len(sound.shape) > 1:
  if sound.shape[1] == 1:

requirements.txt

@@ -1,3 +1,6 @@
+scipy
+numpy
+soundfile
 python-levenshtein
 torch
 torchelastic
@@ -9,3 +12,4 @@ matplotlib
 flask
 sox
 sklearn
+soundfile