def predict(self, path):
self.eval()
if path.find('.wav'):
wav = feature.load_audio(wav_path=path, wav=None)
else:
wav = feature.load_audio(wav=path)
spec = feature.spectrogram(wav)
spec.unsqueeze_(0)
x_lens = spec.size(-1)
out = self.cnn(spec)
out_len = torch.tensor([out.size(-1)])
text = self.decode(out, out_len)
self.train()
return text[0]
def __getitem__(self, index):
wav, transcript = self.idx[index]
wav = feature.load_audio(os.path.join(dataPath, wav))
spect = feature.spectrogram(wav)
transcript = list(
filter(None, [self.labels.get(x) for x in transcript]))
return spect, transcript
def __getitem__(self, index):
wav, transcript = self.idx[index]
wav = feature.load_audio(wav)
spect = feature.spectrogram(wav)
transcript = list(filter(None, [self.vocabulary.get(x) for x in transcript]))
return spect, transcript
def predict(self, path):
self.eval()
wav = feature.load_audio(path)
spec = feature.spectrogram(wav)
spec.unsqueeze_(0)
x_lens = spec.size(-1)
out = self.cnn(spec)
out_len = torch.tensor([out.size(-1)])
text = self.decode(out, out_len)
self.train()
return text[0]
def predict(f):
wav = feature.load_audio(f)
spec = feature.spectrogram(wav)
spec.unsqueeze_(0)
with torch.no_grad():
y = model.cnn(spec)
y = F.softmax(y, 1)
y_len = torch.tensor([y.size(-1)])
y = y.permute(0, 2, 1) # B * T * V
print("decoding")
out, score, offset, out_len = decoder.decode(y, y_len)
return translate(model.vocabulary, out[0][0], out_len[0][0])
def main():
file_path = 'music_data/shortName.flac'
y = f.load_audio(file_path)
# COMPUTE SPECTROGRAM
log_mel_spectrogram = f.compute_spectrogram(y, 2048, 1024, 40)
# COMPUTE ONSET DETECTION FUNCTION
# skip normalization
# norm_spectrogram = o.normalize_frequencies(log_mel_spectrogram)
odf = o.compute_odf(log_mel_spectrogram) #o.compute_odf(norm_spectrogram)
# DETECT ONSETS
peaks = o.apply_threshold(odf, 1500)
print(odf)
# print(peaks)
for i, p in enumerate(peaks):
if p > 0:
print(i/100, ' ', p)
'street_fold{}_train.txt'.format(1))
evaluate_file = os.path.join(evaluation_setup_folder,
'street_fold{}_evaluate.txt'.format(1))
desc_dict = feature.load_desc_file(train_file, __class_labels) #make dict
desc_dict.update(feature.load_desc_file(
evaluate_file,
__class_labels)) # contains labels for all the audio in the dataset
#till here labels are stored for only 1st fold
# Extract features for all audio files, and save it along with labels
for audio_filename in os.listdir(audio_folder):
audio_file = os.path.join(audio_folder, audio_filename)
print('Extracting features and label for : {}'.format(audio_file))
y, sr = feature.load_audio(
audio_file, mono=is_mono,
fs=sr) #y = audio data [shape=(signal_length, channel)]
mbe = None
#now we extract mel band energies for mono or binaural audio
if is_mono:
mbe = feature.extract_mbe(y, sr, nfft, nb_mel_bands).T
else:
for ch in range(y.shape[0]): #for each channel extract mbe
mbe_ch = feature.extract_mbe(y[ch, :], sr, nfft,
nb_mel_bands).T
if mbe is None:
mbe = mbe_ch
else:
mbe = np.concatenate((mbe, mbe_ch), 1)
