def _preprocess_speaker_aishell2(speaker_dir, suffix, out_dir: Path,
skip_existing: bool, hparams, others_params):
speaker_dir, spk_id = speaker_dir
trans_dict = others_params["trans_dict"]
detach_label_and_embed_utt = others_params["detach_label_and_embed_utt"]
metadata = []
wav_fpath_list = speaker_dir.glob("*." + suffix)
utt_fpath_list = list(speaker_dir.glob("*." + suffix))
utt_num = len(utt_fpath_list)
# Iterate over each wav
for wav_fpath in wav_fpath_list:
assert wav_fpath.exists(), str(wav_fpath) + " not exist."
# Process each utterance
wav, _ = librosa.load(str(wav_fpath), hparams.sample_rate)
wav_abs_max = np.max(np.abs(wav))
wav_abs_max = wav_abs_max if wav_abs_max > 0.0 else 1e-8
wav = wav / wav_abs_max * hparams.rescaling_max # norm
# wav_bak = wav
# denoise
if len(wav) > hparams.sample_rate * (0.3 + 0.1):
noise_wav = np.concatenate([
wav[:int(hparams.sample_rate * 0.15)],
wav[-int(hparams.sample_rate * 0.15):]
])
profile = logmmse.profile_noise(noise_wav, hparams.sample_rate)
wav = logmmse.denoise(wav, profile, eta=0)
# trim silence
wav = audio.trim_silence(wav, 30) # top_db: smaller for noisy
# audio.save_wav(wav_bak, str(wav_fpath.name), hparams.sample_rate)
# audio.save_wav(wav, str(wav_fpath.name).replace('.wav','_trimed.wav'),
# hparams.sample_rate)
text = trans_dict[wav_fpath.stem]
# Chinese to Pinyin
pinyin = " ".join(get_pinyin(text, std=True, pb=True))
# print(wav_fpath.name, wav_fpath.stem)
random_uttBasename_forSpkEmbedding = None
if detach_label_and_embed_utt:
random_uttBasename_forSpkEmbedding = utt_fpath_list[
np.random.randint(utt_num)].stem
metadata.append(
process_utterance(wav, pinyin, out_dir, wav_fpath.stem,
skip_existing, hparams,
random_uttBasename_forSpkEmbedding))
return [m for m in metadata if m is not None]
def _preprocess_speaker_SLR38(speaker_dir, suffix, out_dir: Path,
skip_existing: bool, hparams, others_params):
detach_label_and_embed_utt = others_params["detach_label_and_embed_utt"]
wav_fpath_list = speaker_dir.glob("*." + suffix)
text_fpath_list = speaker_dir.glob("*.txt")
metadata = []
# Iterate over each wav
utt_fpath_list = list(speaker_dir.glob("*." + suffix))
utt_num = len(utt_fpath_list)
for wav_fpath, txt_fpath in zip(wav_fpath_list, text_fpath_list):
assert wav_fpath.exists(), str(wav_fpath) + " not exist."
assert txt_fpath.exists(), str(wav_fpath) + " not exist."
# Process each utt
wav, _ = librosa.load(str(wav_fpath), hparams.sample_rate)
wav = wav / np.max(np.abs(wav)) * hparams.rescaling_max
# wav_bak = wav
# denoise
if len(wav) > hparams.sample_rate * (0.3 + 0.1):
noise_wav = np.concatenate([
wav[:int(hparams.sample_rate * 0.15)],
wav[-int(hparams.sample_rate * 0.15):]
])
profile = logmmse.profile_noise(noise_wav, hparams.sample_rate)
wav = logmmse.denoise(wav, profile, eta=0)
# trim silence
wav = audio.trim_silence(wav, 30)
# audio.save_wav(wav_bak, str(wav_fpath.name), hparams.sample_rate)
# audio.save_wav(wav, str(wav_fpath.name).replace('.wav','_trimed.wav'),
# hparams.sample_rate)
# get text
text = txt_fpath.read_text()
# Chinese to Pinyin
pinyin = " ".join(get_pinyin(text, std=True, pb=True))
# print(wav_fpath.name, wav_fpath.stem)
random_uttBasename_forSpkEmbedding = None
if detach_label_and_embed_utt:
random_uttBasename_forSpkEmbedding = utt_fpath_list[
np.random.randint(utt_num)].stem
metadata.append(
process_utterance(wav, pinyin, out_dir, wav_fpath.stem,
skip_existing, hparams,
random_uttBasename_forSpkEmbedding))
return [m for m in metadata if m is not None]
def _preprocess_speaker_SLR68(speaker_dir, suffix, out_dir: Path, skip_existing: bool, hparams, others_params):
trans_dict = others_params["trans_dict"]
metadata = []
wav_fpath_list = speaker_dir.glob("*."+suffix)
# Iterate over each entry in the alignments file
for wav_fpath in wav_fpath_list:
assert wav_fpath.exists(), str(wav_fpath)+" not exist."
# Process each utterance
wav, _ = librosa.load(str(wav_fpath), hparams.sample_rate)
text = trans_dict[wav_fpath.name]["text"]
# Chinese to Pinyin
pinyin = " ".join(get_pinyin(text, std=True, pb=True))
# print(wav_fpath.name, wav_fpath.stem)
metadata.append(process_utterance(wav, pinyin, out_dir, wav_fpath.stem,
skip_existing, hparams))
return [m for m in metadata if m is not None]
def _preprocess_speaker_SLR68(speaker_dir, suffix, out_dir: Path,
skip_existing: bool, hparams, others_params):
trans_dict = others_params["trans_dict"]
metadata = []
wav_fpath_list = speaker_dir.glob("*." + suffix)
# Iterate over each entry in the alignments file
for wav_fpath in wav_fpath_list:
assert wav_fpath.exists(), str(wav_fpath) + " not exist."
# Process each utterance
wav, _ = librosa.load(str(wav_fpath), hparams.sample_rate)
# wav_bak = wav
wav = wav / np.max(np.abs(wav)) * hparams.rescaling_max # norm
# denoise
if len(wav) > hparams.sample_rate * (0.3 + 0.1):
noise_wav = np.concatenate([
wav[:int(hparams.sample_rate * 0.15)],
wav[-int(hparams.sample_rate * 0.15):]
])
profile = logmmse.profile_noise(noise_wav, hparams.sample_rate)
wav = logmmse.denoise(wav, profile, eta=0)
# trim silence
wav = audio.trim_silence(wav, 20) # top_db: smaller for noisy
# audio.save_wav(wav_bak, str(wav_fpath.name), hparams.sample_rate)
# audio.save_wav(wav, str(wav_fpath.name).replace('.wav','_trimed.wav'),
# hparams.sample_rate)
text = trans_dict[wav_fpath.name]["text"]
# Chinese to Pinyin
pinyin = " ".join(get_pinyin(text, std=True, pb=True))
# print(wav_fpath.name, wav_fpath.stem)
metadata.append(
process_utterance(wav, pinyin, out_dir, wav_fpath.stem,
skip_existing, hparams))
return [m for m in metadata if m is not None]
