def generate_infer_batch():
num_frames = self.hparams.segment_length * 100
num_overlap_frames = num_frames * self.hparams.overlap_ratio
for wav_path in tqdm(self.wav_list):
#wav_id = os.path.splitext(os.path.basename(wav_path))[0]
wav_id = self.get_save_path_from_filename(wav_path)
audio, sample_rate = vad_ex.read_wave(wav_path)
vad = webrtcvad.Vad(1)
frames = vad_ex.frame_generator(30, audio, sample_rate)
frames = list(frames)
# following line may cause empty output, too many unvoiced in buffer
segments = vad_ex.vad_collector(sample_rate, 30, 300, vad, frames)
total_wav = b""
for i, segment in enumerate(segments):
total_wav += segment
wav_arr = np.frombuffer(total_wav, dtype=np.int16)
if len(wav_arr) == 0:
#return [],[],False
continue
wav_arr = np.pad(wav_arr, (0, max(0, sample_rate*(self.hparams.segment_length+0.015)-len(wav_arr))), 'constant', constant_values=(0, 0))
logmel_feats = logfbank(wav_arr, samplerate=sample_rate, nfilt=40)
total_len = logmel_feats.shape[0]
num_fix_mels = int((total_len - num_overlap_frames) // (num_frames - num_overlap_frames))
fix_mels = []
for dvec_idx in range(num_fix_mels):
start_idx = int((num_frames - num_overlap_frames) * dvec_idx)
end_idx = int(start_idx + num_frames)
fix_mels.append(logmel_feats[start_idx:end_idx, :])
fix_mels = np.asarray(fix_mels)
self.keys.append(wav_id)
self.fix_mel_lengths.append(len(fix_mels))
for fix_mel in fix_mels:
yield fix_mel
def create_infer_batch(self):
# self.hparams.in_wav1, self.hparams.in_wav2 are full paths of the wav file
# for ex) /home/hdd2tb/ninas96211/dev_wav_set/id10343_pCDWKHjQjso_00002.wav
wavs_list = [self.hparams.in_wav1, self.hparams.in_wav2]
# file_name for ex) id10343_pCDWKHjQjso_00002
for wav_path in wavs_list:
wav_id = os.path.splitext(os.path.basename(wav_path))[0]
audio, sample_rate = vad_ex.read_wave(wav_path)
vad = webrtcvad.Vad(1)
frames = vad_ex.frame_generator(30, audio, sample_rate)
frames = list(frames)
segments = vad_ex.vad_collector(sample_rate, 30, 300, vad, frames)
total_wav = b""
for i, segment in enumerate(segments):
total_wav += segment
print(wav_id + " : " + str(i) + "th segment appended")
# Without writing, unpack total_wav into numpy [N,1] array
# 16bit PCM 기준 dtype=np.int16
wav_arr = np.frombuffer(total_wav, dtype=np.int16)
print("read audio data from byte string. np array of shape:" +
str(wav_arr.shape))
logmel_feats = logfbank(wav_arr, samplerate=sample_rate, nfilt=40)
# file_name for ex, 'id10343_pCDWKHjQjso_00002'
self.save_dict[wav_id] = logmel_feats
num_frames = self.hparams.segment_length * 100
num_overlap_frames = num_frames * self.hparams.overlap_ratio
dvector_dict = {}
match = False
prev_wav_name = ""
for wav_name, feats in self.save_dict.items():
if wav_name.split("_")[0] == prev_wav_name:
print("spk_id" + wav_name.split("_")[0])
match = True
total_len = feats.shape[0]
num_dvectors = int((total_len - num_overlap_frames) //
(num_frames - num_overlap_frames))
print("num dvec:" + str(num_dvectors))
dvectors = []
for dvec_idx in range(num_dvectors):
start_idx = int((num_frames - num_overlap_frames) * dvec_idx)
end_idx = int(start_idx + num_frames)
print("wavname: " + wav_name + " start_idx: " + str(start_idx))
print("wavname: " + wav_name + " end_idx: " + str(end_idx))
dvectors.append(feats[start_idx:end_idx, :])
dvectors = np.asarray(dvectors)
dvector_dict[wav_name] = dvectors
prev_wav_name = wav_name.split("_")[0]
wav1_data = list(dvector_dict.values())[0]
wav2_data = list(dvector_dict.values())[1]
print("match: " + str(match))
print("wav1_data.shape:" + str(wav1_data.shape))
print("wav2_data.shape:" + str(wav2_data.shape))
return wav1_data, wav2_data, match
def vad_process(self, path):
# VAD Process
if self.data_type == "vox1":
audio, sample_rate = vad_ex.read_wave(path)
elif self.data_type == "vox2":
audio, sample_rate = vad_ex.read_m4a(path)
elif "libri" in self.data_type:
audio, sample_rate = vad_ex.read_libri(path)
elif self.data_type == "emotional_actors":
audio, sample_rate = vad_ex.read_wave(path)
elif self.data_type == "darpa_timit":
audio, sample_rate = vad_ex.read_wave(path)
vad = webrtcvad.Vad(1)
frames = vad_ex.frame_generator(30, audio, sample_rate)
frames = list(frames)
segments = vad_ex.vad_collector(sample_rate, 30, 300, vad, frames)
total_wav = b""
for i, segment in enumerate(segments):
total_wav += segment
# Without writing, unpack total_wav into numpy [N,1] array
# 16bit PCM 기준 dtype=np.int16
wav_arr = np.frombuffer(total_wav, dtype=np.int16)
print("read audio data from byte string. np array of shape:" +
str(wav_arr.shape))
return wav_arr, sample_rate
def vad_process(self, path):
wav_id = os.path.splitext(os.path.basename(path))[0]
audio, sample_rate = vad_ex.read_wave(path)
vad = webrtcvad.Vad(1)
frames = vad_ex.frame_generator(30, audio, sample_rate)
frames = list(frames)
segments = vad_ex.vad_collector(sample_rate, 30, 300, vad, frames)
total_wav = b""
for i, segment in enumerate(segments):
total_wav += segment
#print(wav_id + " : " + str(i) + "th segment appended")
# Without writing, unpack total_wav into numpy [N,1] array
# 16bit PCM 기준 dtype=np.int16
wav_arr = np.frombuffer(total_wav, dtype=np.int16)
#print("read audio data from byte string. np array of shape:"+str(wav_arr.shape))
logmel_feats = logfbank(wav_arr, samplerate=sample_rate, nfilt=40)
return logmel_feats
def infer_batch_generator(self):
#import pdb;pdb.set_trace()
batch_size = 640
wavs_list = self.hparams.in_wav1
for wav_path in wavs_list:
# Get voiced wav_arr
if self.hparams.dataset == 'libri':
out_path = self.libri_spkid_outpath(wav_path)
else:
out_path = ''
audio, sample_rate = vad_ex.read_wave(wav_path)
vad = webrtcvad.Vad(1)
frames = vad_ex.frame_generator(30, audio, sample_rate)
frames = list(frames)
segments = vad_ex.vad_collector(sample_rate, 30, 300, vad, frames)
total_wav = b""
for i, segment in enumerate(segments):
total_wav += segment
wav_arr = np.frombuffer(total_wav, dtype=np.int16)
if len(wav_arr) == 0:
continue
# Pad when less than 1.6s
wav_arr = np.pad(wav_arr, (0, max(0, 25840 - len(wav_arr))),
'constant',
constant_values=(0, 0))
# Get logmel
logmel_feats = logfbank(wav_arr, samplerate=sample_rate, nfilt=40)
# Get fixed length log mels
num_frames = self.hparams.segment_length * 100
num_overlap_frames = num_frames * self.hparams.overlap_ratio
total_len = logmel_feats.shape[0]
num_fix_len_mels = int((total_len - num_overlap_frames) //
(num_frames - num_overlap_frames))
fix_len_mels = []
for mel_idx in range(num_fix_len_mels):
start_idx = int((num_frames - num_overlap_frames) * mel_idx)
end_idx = int(start_idx + num_frames)
fix_len_mels.append(logmel_feats[start_idx:end_idx, :])
fix_len_mels = np.asarray(fix_len_mels)
# Queue and pop every 640
for fix_len_mel in fix_len_mels:
last_item = (fix_len_mel, out_path)
self.dq.append(last_item)
self.dq_size += 1
if len(fix_len_mels) == 0:
self.bad_cnt += 1
self.cnt += 1
if self.dq_size >= batch_size:
res = []
for i in range(batch_size):
res.append(self.dq.popleft())
self.dq_size -= 1
yield res
# When remains a lot of 640 clusters
while self.dq_size >= batch_size:
res = []
for i in range(batch_size):
res.append(self.dq.popleft())
self.dq_size -= 1
yield res
# the last remaining cluster, if mod 640 ==0, append dummy 640 logmels
res = []
while self.dq_size > 0:
res.append(self.dq.popleft())
self.dq_size -= 1
remaining = (batch_size - len(res))
last_item = (last_item[0], '')
if remaining != 0:
for i in range(remaining):
res.append(last_item)
yield res
else:
yield res
res = []
for i in range(batch_size):
res.append(last_item)
yield res
