def single_mix(combo_idx, split_list, datapath):
assert len(combo_idx) == len(split_list)
mix_rate = 1.0 / float(len(split_list))
wav_list = []
prefix = 'mix'
mid_name = ''
for part_idx in range(len(split_list)):
idx, path = split_list[part_idx][combo_idx[part_idx]]
wav, _ = librosa.load(path, sr=16000)
wav_list.append(wav)
mid_name += '-%05d' % idx
mix_wav = np.zeros_like(wav_list[0])
for wav in wav_list:
mix_wav += wav * mix_rate #mix two signals
wav_name = prefix + mid_name + '.wav'
wavfile.write('%s/mix_wav/%s' % (datapath, wav_name), 16000, mix_wav)
F_mix = utils.fast_stft(mix_wav)
name = prefix + mid_name + '.npy'
store_path = '%s/mix/%s' % (datapath, name)
np.save(store_path, F_mix)
with open('%s/mix_log.txt' % datapath, 'a') as f:
f.write(name)
f.write('\n')
def single_mix(combo_idx, split_list, database_repo):
assert len(combo_idx) == len(split_list)
mix_rate = 1.0 / float(len(split_list))
wav_list = []
prefix = "mix"
mid_name = ""
for part_idx in range(len(split_list)):
idx, path = split_list[part_idx][combo_idx[part_idx]]
wav, _ = librosa.load(path, sr=16000)
wav_list.append(wav)
mid_name += '-%05d' % idx
# mix wav file
mix_wav = np.zeros_like(wav_list[0])
for wav in wav_list:
mix_wav += wav * mix_rate
# save mix wav file
wav_name = prefix + mid_name + '.wav'
wavfile.write('%s/mix_wav/%s' % (database_repo, wav_name), 16000, mix_wav)
# transfer mix wav to TF domain
F_mix = utils.fast_stft(mix_wav)
name = prefix + mid_name + ".npy"
store_path = '%s/mix/%s' % (database_repo, name)
# save mix as npy file
np.save(store_path, F_mix)
# save mix log
with open('%s/mix_log.txt' % database_repo, 'a') as f:
f.write(name)
f.write("\n")
def preprocess_audio(video_name=VIDEO_NAME, sr=16000):
path = "%s.wav" % video_name
data, _ = librosa.load(path, sr=sr)
data = utils.fast_stft(data)
name = 'preprocessed-%s' % video_name
np.save('%s.npy' % name, data)
def single_audio_to_npy(audio_path_list,
database_repo=DATABASE_REPO_PATH,
fix_sr=16000):
for idx, path in audio_path_list:
data, _ = librosa.load(path, sr=fix_sr)
data = utils.fast_stft(data)
name = 'single-%05d' % idx
with open('%s/single_TF.txt' % database_repo, 'a') as f:
f.write('%s.npy' % name)
f.write('\n')
np.save(('%s/single/%s.npy' % (database_repo, name)), data)
def single_audio_to_npy(audio_path_list, database_repo=DATABASE_REPO_PATH, fix_sr=16000):
for idx, path in audio_path_list:
print('\rsingle npy generating... %d' % ((idx / len(audio_path_list)) * 100), end='')
data, _ = librosa.load(path, sr=fix_sr)
data = utils.fast_stft(data, power=True)
name = 'single-%06d' % idx
with open('%s/single_TF.txt' % database_repo, 'a') as f:
f.write('%s.npy' % name)
f.write('\n')
np.save(('%s/single/%s.npy' % (database_repo, name)), data)
print()
def audio_to_numpy(audio_path_list, data_path=database_path, fix_sr=16000):
for idx, path in audio_path_list:
print('\r aduio numpy generating... %d' %
((idx / len(audio_path_list)) * 100),
end='')
data, _ = librosa.load(path, sr=fix_sr)
data = utils.fast_stft(data)
name = 'single-%05d' % idx
with open('%s/single_TF.txt' % data_path, 'a') as f:
f.write('%s.npy' % name)
f.write('\n')
np.save(('%s/single/%s.npy' % (data_path, name)), data)
print()
# options
SHAPE_CHECK = 0
MODEL_CHECK1 = 0
MODEL_CHECK2 = 1
data1, sr1 = librosa.load('../../data/audio/audio_train/trim_audio_train0.wav',
sr=16000)
data2, sr2 = librosa.load('../../data/audio/audio_train/trim_audio_train1.wav',
sr=16000)
mix = data1 * 0.5 + data2 * 0.5
if SHAPE_CHECK:
# check shape
print(data1.shape)
D1 = avp.fast_stft(data1)
print(D1.shape)
T1 = avp.fast_istft(D1)
print(T1.shape)
# check model
if MODEL_CHECK1:
F1 = avp.fast_stft(data1)
F1 = np.expand_dims(F1, axis=0)
F2 = avp.fast_stft(data2)
F2 = np.expand_dims(F2, axis=0)
FM = avp.fast_stft(mix)
FM = np.expand_dims(FM, axis=0)
cRM1 = np.abs(F1) / np.abs(FM)
cRM1[~np.isfinite(cRM1)] = 0
def generate_mix_sample(audio_path_list,num_speaker,fix_sr=16000,verbose=0):
'''
generate mix sample from audios in the list
:param audio_path_list: list contains path of the wav audio file
:param num_speaker: specify the task for speech separation
:param fix_sr: fix sample rate
'''
# initiate variables
# shape of F_mix = (298,257,2)
# shpae of crm = (298,257,2)
data_list = []
F_list = [] # STFT list for each sample
cRM_list = []
mix_name = "mix"
crm_name = "crm"
post_name = ""
# import data
for i in range(num_speaker):
idx,path =audio_path_list[i]
post_name += "-%05d"%idx
data, _ = librosa.load(path,sr=fix_sr)
data_list.append(data)
# create mix audio according to mix rate
mix_rate = 1.0 / float(num_speaker)
mix = np.zeros(shape=data_list[0].shape)
for data in data_list:
mix += data*mix_rate
# transfrom data via STFT and several preprocessing function
for i in range(num_speaker):
F = utils.fast_stft(data_list[i],power=False)
F_list.append(F)
F_mix = utils.fast_stft(mix,power=False)
# create cRM for each speaker and fill into y_sample
for i in range(num_speaker):
cRM_list.append(utils.fast_cRM(F_list[i],F_mix))
# return values
if verbose == 1:
print('shape of X: ',F_mix.shape)
for i in range(len(cRM_list)):
print('shape of cRM%s :'%i,cRM_list[i].shape)
# save record in txt
mix_name += post_name
crm_name += post_name
# write txt
with open('audio_database/dataset.txt','a') as f:
f.write(mix_name+".npy")
for i in range(len(cRM_list)):
line = " " + crm_name + ("-%05d"%audio_path_list[i][0]) + ".npy"
f.write(line)
f.write("\n")
# save file as npy
np.save(('audio_database/mix/%s.npy'%mix_name), F_mix)
for i in range(len(cRM_list)):
name = crm_name + ("-%05d"%audio_path_list[i][0])
np.save(('audio_database/crm/%s.npy'%name), cRM_list[i])
def single_audio_to_npy(audio_path_list,fix_sr=16000):
for idx,path in audio_path_list:
data, _ = librosa.load(path, sr=fix_sr)
data = utils.fast_stft(data)
name = 'single-%05d'%idx
np.save(('audio_database/single/%s.npy'%name),data)
