def spec_feature_extraction(wav_file):
'''
Extracts the mel and magnitude spectrogram from the given audio file
:param wav_file: The audio file to extract spectrogram from
returns the mel and magnitude spectrogram of the given audio file
'''
mel, mag = get_spectrograms(wav_file)
return mel, mag
def make_one_dataset(filename,total,display=False):
global finish
speaker_id = filename.strip().split('/')[-1][:-4]
mel_spec, lin_spec = get_spectrograms(filename)
# wav = preprocess_wav(Path(filename))
# d_mel, d_mel_slices = d_wav2spec(wav)
print('[Processor] - processing {}/{} {} | mel: {} '.format(
finish*WORKERS, total, speaker_id, mel_spec.shape), end='\r')
result = {}
result['speaker_id'] = speaker_id
result['mel_spec'] = mel_spec
result['lin_spec'] = lin_spec
finish += 1
return result
def make_one_dataset(filename,total,display=False):
global finish
sub_filename = filename.strip().split('/')[-1]
groups = re.match(r'(\d+)_(\d+)_(\d+)_(\d+)\.wav', sub_filename).groups() # format: p{speaker}_{sid}.wav
speaker_id = groups[0]
utt_id = '_'.join(groups[1:])
mel_spec, lin_spec = get_spectrograms(filename)
wav = preprocess_wav(Path(filename))
d_mel, d_mel_slices = d_wav2spec(wav)
print('[Processor] - processing {}/{} s{}-{} | mel: {} | d_mel: {}'.format(
finish*WORKERS, total, speaker_id, utt_id, mel_spec.shape, d_mel.shape), end='\r')
result = {}
result['speaker_id'] = speaker_id
result['utt_id'] = utt_id
result['d_mel_spec'] = d_mel
result['d_mel_slices'] = d_mel_slices
result['mel_spec'] = mel_spec
result['lin_spec'] = lin_spec
finish += 1
return result
def save_mel_spectragram(wav_file, npy_dir):
mel, _ = get_spectrograms(wav_file)
_, name = os.path.split(wav_file)
stem, ext = os.path.splitext(name)
np.save(os.path.join(npy_dir, stem + '.npy'), mel)
def spec_feature_extraction(wav_file):
mel, mag = get_spectrograms(wav_file)
return mel, mag
def audio2feature():
data_list = []
label_list = []
audio_list = []
# read file
#files = glob.glob(config['dataset_path']+'/*/*.mp3')
paths = glob.glob(config['dataset_path'] + '/*')
for p in paths: ####song
print('song_name:' + p)
files = glob.glob(p + '/*.mp3')
for filename in files:
index1 = filename.find('\\') + 1 ####../dataset/train\01
index2 = filename.find('\\', index1)
song_label = p[index1:index2]
#print('song label:{}'.format(song_label)) ###song label
print(filename)
if filename.find('solo') != -1:
instru_class = filename[-11:-9] #class label
else:
instru_class = filename[-6:-4]
if int(instru_class) >= config['class']:
print('dump %s cause it is class %s' %
(filename, instru_class))
continue
y, sr = librosa.core.load(
filename, offset=0.0, sr=config['sr']
) # read after 1 seconds #mono=True(convert signal to mono)
#y, sr = librosa.load(filename, sr=config['sr'])
Len = y.shape[0]
count = 0 #segment count
st_idx = 0
end_idx = st_idx + config['audio_samples_frame_size']
next_idx = st_idx + config['audio_samples_hop_length']
while st_idx < Len:
#### label
label = [song_label, instru_class, count]
label_list.append(label)
if end_idx > Len:
end_idx = Len ####last is too short?
print(label) #[song,instrument,count]
#### audio
audio = np.zeros(config['audio_samples_frame_size'],
dtype='float32')
audio[:end_idx - st_idx] = y[st_idx:end_idx]
if config['pre'] == 'from hung-yi':
feature, _ = get_spectrograms(audio)
else:
feature = get_melspec(audio, config)
data_list.append(feature)
audio_list.append(audio)
'''
# output audio
outputname=os.path.join('cut{:03d}.wav'.format(count))
print(outputname)
librosa.output.write_wav(outputname,audio,config['sr'])
input()
'''
count += 1
st_idx = next_idx
end_idx = st_idx + config['audio_samples_frame_size']
next_idx = st_idx + config['audio_samples_hop_length']
###### normal
if config['nor'] == 'yes':
print('starting normalize')
if (config['dataset_path'][-4:] == 'test'):
with open(os.path.join(config['attr_path'], 'attr.pkl'),
'rb') as f:
attr = pickle.load(f)
mean = attr['mean']
std = attr['std']
print('mean, std:{},{}'.format(mean, std))
else:
data_temp = np.concatenate(data_list)
mean = np.mean(data_temp, axis=0)
std = np.std(data_temp, axis=0)
'''
data_temp shape:(48440, 128)
mean shape:(128,)
'''
attr = {'mean': mean, 'std': std}
with open(os.path.join(config['npy_path'], 'attr.pkl'), 'wb') as f:
pickle.dump(attr, f)
##### normalize!!!!!!
data_list = (data_list - mean) / std
if config['pre'] == 'from hung-yi':
print('hung-yi transpose')
temp = []
for val in data_list:
val = val.T #numpy T /// pytorch transpose
temp.append(val)
data_list = temp
# save
data_name = os.path.join(config['npy_path'], config['data_npy'])
label_name = os.path.join(config['npy_path'], config['label_npy'])
np.save(data_name, data_list)
np.save(label_name, label_list)
### to wave
audio_name = os.path.join(config['npy_path'], config['audio_npy'])
np.save(audio_name, audio_list)
h5py_path = sys.argv[1]
filename_groups = defaultdict(lambda: [])
with h5py.File(h5py_path, 'w') as f_h5:
grps = [f_h5.create_group('train'), f_h5.create_group('test')]
filenames = sorted(glob.glob(os.path.join(root_dir, '*/*/*.flac')))
for filename in filenames:
# divide into groups
speaker_id, chapter_id, segment_id = filename.strip().split(
'/')[-1].strip('.flac').split('-')
filename_groups[speaker_id].append(filename)
for speaker_id, filenames in filename_groups.items():
print('processing {}'.format(speaker_id))
for filename in filenames[:-1]:
print(filename)
speaker_id, chapter_id, segment_id = filename.strip().split(
'/')[-1].strip('.flac').split('-')
mel_spec, lin_spec = get_spectrograms(filename)
grps[0].create_dataset('{}/{}-{}/mel'.format(speaker_id, chapter_id, segment_id), \
data=mel_spec, dtype=np.float32)
grps[0].create_dataset('{}/{}-{}/lin'.format(speaker_id, chapter_id, segment_id),\
data=lin_spec, dtype=np.float32)
# the last segment put into testset
filename = filenames[-1]
speaker_id, chapter_id, segment_id = filename.strip().split(
'/')[-1].strip('.flac').split('-')
mel_spec, lin_spec = get_spectrograms(filename)
grps[1].create_dataset('{}/{}-{}/mel'.format(speaker_id, chapter_id, segment_id), \
data=mel_spec, dtype=np.float32)
grps[1].create_dataset('{}/{}-{}/lin'.format(speaker_id, chapter_id, segment_id), \
data=lin_spec, dtype=np.float32)