def get_random_wav_and_label(self, tar_wavfiles, ntar_wavfiles):
"""
:return: wav: raw wave. float32. shape=(t, ),
label: 1 if target, 0 otherwise. int32.
melspec: mel-spectrogram. float32. shape=(t, n_mels)
"""
wavfiles, label = (
tar_wavfiles,
self.tar_labels) if np.random.sample(1) <= self.tar_ratio else (
ntar_wavfiles, self.ntar_labels)
wavfile = wavfiles[np.random.randint(0, len(wavfiles))]
if type(wavfile) == bytes:
wavfile = wavfile.decode()
if wavfile.endswith('arr'): # pyarrow format
wav = read_wav_from_arr(wavfile)
else:
wav = read_wav(wavfile, sr=hp.signal.sr)
wav = trim_wav(wav)
wav = crop_random_wav(wav, self.length)
wav = augment_volume(wav)
wav = fix_length(wav, self.length) # padding
melspec = wav2melspec_db(wav,
sr=hp.signal.sr,
n_fft=hp.signal.n_fft,
win_length=hp.signal.win_length,
hop_length=hp.signal.hop_length,
n_mels=hp.signal.n_mels,
min_db=hp.signal.min_db,
max_db=hp.signal.max_db)
melspec = np.float32(melspec)
label = np.float32(label)
return wav, melspec, label
def _get_wav_and_melspec(wav_file, length, is_training=True):
'''
the range of values of wav is [-1, 1].
'''
wav = read_wav(wav_file, sr=hp.signal.sr)
wav = trim_wav(wav)
# divide wav into chunks that have the given length and one is randomly selected in training, but first chunk in generation.
n_clips = math.ceil(len(wav) / length) if is_training else 1
idx = random.randrange(n_clips)
start, end = length * idx, length * (idx + 1)
wav = wav[start:end]
assert (len(wav) <= length)
wav = fix_length(wav, length) # padding in case of last chunk.
melspec = wav2melspec_db(wav,
sr=hp.signal.sr,
n_fft=hp.signal.n_fft,
win_length=hp.signal.win_length,
hop_length=hp.signal.hop_length,
n_mels=hp.signal.n_mels,
min_db=hp.signal.min_db,
max_db=hp.signal.max_db)
wav = np.expand_dims(wav, -1)
return wav, melspec.astype(np.float32)
def _load_random_wav(self, speaker_id):
wavfile = self.audio_meta.get_random_audio(speaker_id)
wav = read_wav(wavfile, hp.signal.sr)
# wav = trim_wav(wav)
length = int(hp.signal.duration * hp.signal.sr)
wav = crop_random_wav(wav, length=length)
wav = fix_length(wav, length, mode='reflect')
return wav # (t, n_mel)
def get_random_wav(self, wavfile):
"""
:param: wavfile: a raw wave file.
:return: wav: raw wave. float32. shape=(t, ),
melspec: mel-spectrogram. float32. shape=(t, n_mels),
wavfile: the raw wave file.
"""
wav = read_wav(wavfile, sr=hp.signal.sr)
wav = trim_wav(wav)
wav = fix_length(wav, self.length) # crop from the beginning.
melspec = wav2melspec_db(wav,
sr=hp.signal.sr,
n_fft=hp.signal.n_fft,
win_length=hp.signal.win_length,
hop_length=hp.signal.hop_length,
n_mels=hp.signal.n_mels,
min_db=hp.signal.min_db,
max_db=hp.signal.max_db)
melspec = np.float32(melspec)
return wav, melspec, wavfile
def _get_wav_and_melspec(wav_file, length=None, is_training=True):
wav = read_wav(wav_file, sr=hp.signal.sr)
wav = trim_wav(wav)
if length:
n_clips = math.ceil(len(wav) / length) if is_training else 1
idx = random.randrange(n_clips)
start, end = length * idx, length * (idx + 1)
wav = wav[start:end]
assert (len(wav) <= length)
wav = fix_length(wav, length) # padding
melspec = wav2melspec_db(wav,
sr=hp.signal.sr,
n_fft=hp.signal.n_fft,
win_length=hp.signal.win_length,
hop_length=hp.signal.hop_length,
n_mels=hp.signal.n_mels,
min_db=hp.signal.min_db,
max_db=hp.signal.max_db)
wav = np.expand_dims(wav, -1)
return wav, melspec.astype(np.float32)
ckpt = args.ckpt if args.ckpt else tf.train.latest_checkpoint(hp.logdir)
pred_conf = PredictConfig(
model=model,
input_names=['x'],
output_names=['embedding/embedding', 'prediction'],
session_init=SaverRestore(ckpt) if ckpt else None)
embedding_pred = OfflinePredictor(pred_conf)
embedding, pred_speaker_id = embedding_pred(mel_spec)
# get a random audio of the predicted speaker.
wavfile_pred_speaker = np.array(map(lambda s: audio_meta_train.get_random_audio(s), pred_speaker_id))
length = int(hp.signal.duration * hp.signal.sr)
wav_pred_speaker = np.array(
map(lambda w: fix_length(read_wav(w, hp.signal.sr, duration=hp.signal.duration), length),
wavfile_pred_speaker))
# write audio
tf.summary.audio('wav', wav, hp.signal.sr, max_outputs=10)
tf.summary.audio('wav_pred', wav_pred_speaker, hp.signal.sr, max_outputs=10)
# write prediction
speaker_name = [audio_meta.speaker_dict[sid] for sid in speaker_id]
pred_speaker_name = [audio_meta_train.speaker_dict[sid] for sid in pred_speaker_id]
meta = [tuple(audio_meta.meta_dict[sid][k] for k in audio_meta.target_meta_field()) for sid in speaker_id] if hp.embed.meta_path else None
pred_meta = [tuple(audio_meta_train.meta_dict[sid][k] for k in audio_meta_train.target_meta_field()) for sid in pred_speaker_id] if hp.train.meta_path else None
prediction = ['{} ({}) -> {} ({})'.format(s, s_meta, p, p_meta)
for s, p, s_meta, p_meta in zip(speaker_name, pred_speaker_name, meta, pred_meta)]
tf.summary.text('prediction', tf.convert_to_tensor(prediction))
output_names=['embedding/embedding', 'prediction'],
session_init=SaverRestore(ckpt) if ckpt else None,
)
embedding_pred = OfflinePredictor(pred_conf)
embedding, pred_speaker_id = embedding_pred(mel_spec)
# get a random audio of the predicted speaker.
wavfile_pred_speaker = np.array(
map(lambda s: audio_meta_train.get_random_audio(s), pred_speaker_id))
length = int(hp.signal.duration * hp.signal.sr)
wav_pred_speaker = np.array(
map(
lambda w: fix_length(
read_wav(w, hp.signal.sr, duration=hp.signal.duration), length
), wavfile_pred_speaker))
# write audio
tf.summary.audio('wav', wav, hp.signal.sr, max_outputs=10)
tf.summary.audio('wav_pred',
wav_pred_speaker,
hp.signal.sr,
max_outputs=10)
# write prediction
speaker_name = [audio_meta.speaker_dict[sid] for sid in speaker_id]
pred_speaker_name = [
audio_meta_train.speaker_dict[sid] for sid in pred_speaker_id
]