def run_eval(args, checkpoint_path, output_dir, hparams, sentences):
eval_dir = os.path.join(output_dir, 'eval')
log_dir = os.path.join(output_dir, 'logs-eval')
if args.model in ('Both', 'Tacotron-2'):
assert os.path.normpath(eval_dir) == os.path.normpath(
args.mels_dir) # mels_dir = wavenet_input_dir
# Create output path if it doesn't exist
os.makedirs(eval_dir, exist_ok=True)
os.makedirs(log_dir, exist_ok=True)
os.makedirs(os.path.join(log_dir, 'wavs'), exist_ok=True)
os.makedirs(os.path.join(log_dir, 'plots'), exist_ok=True)
log(hparams_debug_string())
synth = Synthesizer()
synth.load(checkpoint_path, hparams)
with open(os.path.join(eval_dir, 'map.txt'), 'w') as file:
for i, text in enumerate(tqdm(sentences)):
if is_korean_text(text):
text = normalize_number(text)
# 한글을 자소 단위로 쪼갠다.
text = split_to_jamo(text, hparams.cleaners)
mel_filename = synth.synthesize(text, i + 1, eval_dir, log_dir,
None)
file.write('{}|{}\n'.format(text, mel_filename))
log('synthesized mel spectrograms at {}'.format(eval_dir))
return eval_dir
def view_method():
text = request.args.get('text')
if is_korean_text(text):
text = normalize_number(text)
text = split_to_jamo(text, hparams.cleaners)
speaker_id = int(request.args.get('speaker_id'))
if text:
return generate_audio_response(text, speaker_id)
else:
return jsonify(success=True), 200
def _get_next_example(self):
"""Gets a single example (input, mel_target, token_target, linear_target, mel_length) from_ disk
"""
if self._train_offset >= len(self._train_meta):
self._train_offset = 0
np.random.shuffle(self._train_meta)
meta = self._train_meta[self._train_offset]
self._train_offset += 1
speaker_id = int(meta[5])
text = meta[6]
if is_korean_text(text):
text = normalize_number(text)
# 한글을 자소 단위로 쪼갠다.
text = split_to_jamo(text, self._cleaner_names)
input_data = np.asarray(text_to_sequence(text, self._cleaner_names), dtype=np.int32)
mel_target = np.load(os.path.join(self._mel_dir, meta[1]))
# Create parallel sequences containing zeros to represent a non finished sequence
token_target = np.asarray([0.] * (len(mel_target) - 1))
linear_target = np.load(os.path.join(self._linear_dir, meta[2]))
return (input_data, mel_target, token_target, linear_target, speaker_id, len(mel_target))
def synthesize(self, text, index, out_dir, log_dir, mel_filename, speaker_id):
hparams = self._hparams
cleaner_names = [x.strip() for x in hparams.cleaners.split(',')]
if is_korean_text(text):
text = normalize_number(text)
text = split_to_jamo(text, cleaner_names)
seq = text_to_sequence(text, cleaner_names)
feed_dict = {
self.model.inputs: [np.asarray(seq, dtype=np.int32)],
self.model.input_lengths: np.asarray([len(seq)], dtype=np.int32),
self.model.speaker_ids: np.asarray([speaker_id], dtype=np.int32)
}
if self.gta:
feed_dict[self.model.mel_targets] = np.load(mel_filename).reshape(1, -1, 80)
if self.gta or not hparams.predict_linear:
mels, alignment = self.session.run([self.mel_outputs, self.alignment], feed_dict=feed_dict)
else:
linear, mels, alignment = self.session.run([self.linear_outputs, self.mel_outputs, self.alignment],
feed_dict=feed_dict)
linear = linear.reshape(-1, hparams.num_freq)
mels = mels.reshape(-1, hparams.num_mels) # Thanks to @imdatsolak for pointing this out
if index is None:
# Generate wav and read it
wav = audio.inv_mel_spectrogram(mels.T, hparams)
audio.save_wav(wav, 'temp.wav', sr=hparams.sample_rate) # Find a better way
chunk = 512
f = wave.open('temp.wav', 'rb')
p = pyaudio.PyAudio()
stream = p.open(format=p.get_format_from_width(f.getsampwidth()),
channels=f.getnchannels(),
rate=f.getframerate(),
output=True)
data = f.readframes(chunk)
while data:
stream.write(data)
data = f.readframes(chunk)
stream.stop_stream()
stream.close()
p.terminate()
return
# Write the spectrogram to disk
# Note: outputs mel-spectrogram files and target ones have same names, just different folders
mel_filename = os.path.join(out_dir, 'speech-mel-{:05d}.npy'.format(index))
np.save(mel_filename, mels, allow_pickle=False)
if log_dir is not None:
# save wav (mel -> wav)
wav = audio.inv_mel_spectrogram(mels.T, hparams)
audio.save_wav(wav, os.path.join(log_dir, 'wavs/speech-wav-{:05d}-mel.wav'.format(index)),
sr=hparams.sample_rate)
if hparams.predict_linear:
# save wav (linear -> wav)
wav = audio.inv_linear_spectrogram(linear.T, hparams)
audio.save_wav(wav, os.path.join(log_dir, 'wavs/speech-wav-{:05d}-linear.wav'.format(index)),
sr=hparams.sample_rate)
if is_korean_char(text):
text = j2h(text)
# save alignments
plot.plot_alignment(alignment, os.path.join(log_dir, 'plots/speech-alignment-{:05d}.png'.format(index)),
info='{}'.format(text), split_title=True)
# save mel spectrogram plot
plot.plot_spectrogram(mels, os.path.join(log_dir, 'plots/speech-mel-{:05d}.png'.format(index)),
info='{}'.format(text), split_title=True)
return mel_filename