def save_spectrogram_tdsv():
""" Select text specific utterance and perform STFT with the audio file.
Audio spectrogram files are divided as train set and test set and saved as numpy file.
Need : utterance data set (VTCK)
"""
print("start text dependent utterance selection")
os.makedirs(config.train_path, exist_ok=True) # make folder to save train file
os.makedirs(config.test_path, exist_ok=True) # make folder to save test file
# =============================================================================================2020/03/23 23:25
# ============================
total_speaker_num = 0
for audio_path in audio_paths:
total_speaker_num += len(os.listdir(audio_path))
train_speaker_num = (total_speaker_num // 10) * 9 # split total data 90% train and 10% test
print("total speaker number : %d" % total_speaker_num)
print("train : %d, test : %d" % (train_speaker_num, total_speaker_num - train_speaker_num))
i = 0
# ============================
# =============================================================================================2020/03/23 23:25
for audio_path in audio_paths:
for folder in os.listdir(audio_path):
speaker_path = os.path.join(audio_path, folder) # path of each speaker
print("%dth speaker processing..." % i)
utterances_spec = []
for utter_name in os.listdir(speaker_path):
utter_path = os.path.join(speaker_path, utter_name) # path of each utterance
# utter_path= os.path.join(audio_path, folder, os.listdir(os.path.join(audio_path, folder))[0])
# if os.path.splitext(os.path.basename(utter_path))[0][-3:] != '001': # if the text utterance doesn't exist pass
# print(os.path.basename(utter_path)[:4], "001 file doesn't exist")
# continue
try:
utter, sr = librosa.core.load(utter_path, config.sr) # load the utterance audio
utter_trim, index = librosa.effects.trim(utter, top_db=14) # trim the beginning and end blank
if utter_trim.shape[0]/sr <= config.hop*(config.tdsv_frame+2): # if trimmed file is too short, then pass
print(os.path.basename(utter_path), "voice trim fail")
continue
S = librosa.core.stft(y=utter_trim, n_fft=config.nfft,
win_length=int(config.window * sr), hop_length=int(config.hop * sr)) # perform STFT
S = keyword_spot(S) # keyword spot (for now, just slice last 80 frames which contains "Call Stella")
utterances_spec.append(S) # make spectrograms list
except:
continue
utterances_spec = np.array(utterances_spec) # list to numpy array
# np.random.shuffle(utterances_spec) # shuffle spectrogram (by person)
# total_num = utterances_spec.shape[0]
# train_num = (total_num//10)*9 # split total data 90% train and 10% test
# print("selection is end")
# print("total utterances number : %d"%total_num, ", shape : ", utterances_spec.shape)
# print("train : %d, test : %d"%(train_num, total_num- train_num))
print(utterances_spec.shape)
if i < train_speaker_num: # save spectrogram as numpy file
np.save(os.path.join(config.train_path, "speaker%d.npy" % i), utterances_spec)
else:
np.save(os.path.join(config.test_path, "speaker%d.npy" % (i - train_speaker_num)), utterances_spec)
i+=1
def save_spectrogram_tdsv():
""" Select text specific utterance and perform STFT with the audio file.
Audio spectrogram files are divided as train set and test set and saved as numpy file.
Need : utterance data set (VTCK)
"""
print("start text dependent utterance selection")
os.makedirs(config.train_path,
exist_ok=True) # make folder to save train file
os.makedirs(config.test_path,
exist_ok=True) # make folder to save test file
utterances_spec = []
for folder in os.listdir(audio_path):
spk_dir = os.path.join(audio_path, folder)
spk_utts = os.listdir(spk_dir)
spk_utts.sort()
utter_path = os.path.join(spk_dir, spk_utts[0])
if os.path.splitext(
os.path.basename(utter_path)
)[0][-3:] != '001': # if the text utterance doesn't exist pass
print(os.path.basename(utter_path)[:4], "001 file doesn't exist")
continue
utter, sr = librosa.core.load(utter_path,
config.sr) # load the utterance audio
utter_trim, index = librosa.effects.trim(
utter, top_db=14) # trim the beginning and end blank
if utter_trim.shape[0] / sr <= config.hop * (
config.tdsv_frame +
2): # if trimmed file is too short, then pass
print(os.path.basename(utter_path), "voice trim fail")
continue
S = librosa.core.stft(y=utter_trim,
n_fft=config.nfft,
win_length=int(config.window * sr),
hop_length=int(config.hop * sr)) # perform STFT
S = keyword_spot(
S
) # keyword spot (for now, just slice last 80 frames which contains "Call Stella")
utterances_spec.append(S) # make spectrograms list
utterances_spec = np.array(utterances_spec) # list to numpy array
np.random.shuffle(utterances_spec) # shuffle spectrogram (by person)
total_num = utterances_spec.shape[0]
train_num = (total_num //
10) * 9 # split total data 90% train and 10% test
print("selection is end")
print("total utterances number : %d" % total_num, ", shape : ",
utterances_spec.shape)
print("train : %d, test : %d" % (train_num, total_num - train_num))
np.save(os.path.join(config.train_path, "train.npy"),
utterances_spec[:train_num]) # save spectrogram as numpy file
np.save(os.path.join(config.test_path, "test.npy"),
utterances_spec[train_num:])
def save_spectrogram_tdsv(path, data_type):
""" Select text specific utterance and perform STFT with the audio file.
Audio spectrogram files are divided as train set and test set and saved as numpy file.
Need : utterance data set (VTCK)
"""
print('Preprocess ' + data_type)
utterances_spec = []
for folder in os.listdir(path):
if not os.path.isdir(os.path.join(path, folder)):
continue
audios = os.listdir(os.path.join(path, folder))
audios.sort()
utter_path = os.path.join(path, folder, audios[0])
if config.train and os.path.splitext(
os.path.basename(utter_path)
)[0][-3:] != '001': # if the text utterance doesn't exist pass
print(os.path.basename(utter_path)[:4], "001 file doesn't exist")
continue
utter, sr = librosa.core.load(utter_path,
config.sr) # load the utterance audio
utter_trim, index = librosa.effects.trim(
utter, top_db=14) # trim the beginning and end blank
if utter_trim.shape[0] / sr <= config.hop * (
config.tdsv_frame +
2): # if trimmed file is too short, then pass
print(os.path.basename(utter_path), "voice trim fail")
continue
S = librosa.core.stft(y=utter_trim,
n_fft=config.nfft,
win_length=int(config.window * sr),
hop_length=int(config.hop * sr)) # perform STFT
S = keyword_spot(
S
) # keyword spot (for now, just slice last 80 frames which contains "Call Stella")
utterances_spec.append(S) # make spectrograms list
utterances_spec = np.array(utterances_spec) # list to numpy array
np.random.shuffle(utterances_spec) # shuffle spectrogram (by person)
total_num = utterances_spec.shape[0]
print("Speaker number : %d" % total_num, ", shape : ",
utterances_spec.shape)
np.save(os.path.join(path, data_type + ".npy"),
utterances_spec) # save spectrogram as numpy file