def get_mel(self, filename):
if not self.load_mel_from_disk:
audio, sampling_rate = load_wav_to_torch(filename)
if sampling_rate != self.stft.sampling_rate:
raise ValueError("{} {} SR doesn't match target {} SR".format(
sampling_rate, self.stft.sampling_rate))
audio_norm = audio / self.max_wav_value
audio_norm = audio_norm.unsqueeze(0)
audio_norm = torch.autograd.Variable(audio_norm, requires_grad=False)
melspec = self.stft.mel_spectrogram(audio_norm)
melspec = torch.squeeze(melspec, 0)
else:
#melspec = torch.from_numpy(np.load(filename))
melspec=torch.load(filename)
assert melspec.size(0) == self.stft.n_mel_channels, (
'Mel dimension mismatch: given {}, expected {}'.format(
melspec.size(0), self.stft.n_mel_channels))
return melspec
def get_mel_audio_pair(self, filename):
audio, sampling_rate = load_wav_to_torch(filename, sr = self.sampling_rate)
if sampling_rate != self.stft.sampling_rate:
raise ValueError(f"{sampling_rate} SR doesn't match target {self.stft.sampling_rate} SR")
# Take segment
if audio.size(0) >= self.segment_length:
max_audio_start = audio.size(0) - self.segment_length
audio_start = random.randint(0, max_audio_start)
audio = audio[audio_start:audio_start+self.segment_length]
else:
audio = torch.nn.functional.pad(
audio, (0, self.segment_length - audio.size(0)), 'constant').data
audio = audio / self.max_wav_value
audio_norm = audio.unsqueeze(0)
audio_norm = torch.autograd.Variable(audio_norm, requires_grad=False)
melspec = self.stft.mel_spectrogram(audio_norm)
melspec = melspec.squeeze(0)
return (melspec, audio, len(audio))
def main():
"""
Launches text to speech (inference).
Inference is executed on a single GPU.
"""
parser = argparse.ArgumentParser(
description='PyTorch Tacotron 2 Inference')
parser = parse_training_args(parser)
args, _ = parser.parse_known_args()
LOGGER.set_model_name("Tacotron2_PyT")
LOGGER.set_backends([
dllg.StdOutBackend(log_file=None,
logging_scope=dllg.TRAIN_ITER_SCOPE,
iteration_interval=1),
dllg.JsonBackend(log_file=args.log_file,
logging_scope=dllg.TRAIN_ITER_SCOPE,
iteration_interval=1)
])
LOGGER.register_metric("tacotron2_frames_per_sec",
metric_scope=dllg.TRAIN_ITER_SCOPE)
LOGGER.register_metric("tacotron2_latency",
metric_scope=dllg.TRAIN_ITER_SCOPE)
LOGGER.register_metric("latency", metric_scope=dllg.TRAIN_ITER_SCOPE)
model, args = load_and_setup_model(parser, args)
log_hardware()
log_args(args)
os.makedirs(args.output_dir, exist_ok=True)
LOGGER.iteration_start()
measurements = {}
anchor_dirs = [
os.path.join(args.dataset_path, anchor)
for anchor in args.training_anchor_dirs
]
metadatas = [load_metadata(anchor) for anchor in anchor_dirs]
stft = TacotronSTFT(args.filter_length, args.hop_length, args.win_length,
args.n_mel_channels, args.sampling_rate, args.mel_fmin,
args.mel_fmax)
with torch.no_grad(), MeasureTime(measurements, "tacotron2_time"):
for speaker_id in range(len(anchor_dirs)):
metadata = metadatas[speaker_id]
for npy_path, text in tqdm(metadata):
seq = text_to_sequence(text, speaker_id, ['basic_cleaners'])
seqs = torch.from_numpy(np.stack(seq)).unsqueeze(0)
seq_lens = torch.IntTensor([len(text)])
wav = load_wav_to_torch(npy_path)
mel = stft.mel_spectrogram(wav.unsqueeze(0))
mel = mel.squeeze()
max_target_len = mel.size(1) - 1
max_target_len += args.n_frames_per_step - max_target_len % args.n_frames_per_step
padded_mel = np.pad(mel, [(0, 0),
(0, max_target_len - mel.size(1))],
mode='constant',
constant_values=args.mel_pad_val)
target = padded_mel[:, ::args.n_frames_per_step]
targets = torch.from_numpy(np.stack(target)).unsqueeze(0)
target_lengths = torch.IntTensor([target.shape[1]])
outputs = model.infer(
to_gpu(seqs).long(),
to_gpu(seq_lens).int(),
to_gpu(targets).half(),
to_gpu(target_lengths).int())
_, mel_out, _, _ = [
output.cpu() for output in outputs if output is not None
]
mel_out = mel_out.squeeze()[:, :mel.size(-1) - 1]
assert (mel_out.shape[-1] == wav.shape[-1] // args.hop_length)
fname = os.path.basename(npy_path)
np.save(os.path.join(args.output_dir, fname),
mel_out,
allow_pickle=False)
# GTA synthesis
# magnitudes = stft.inv_mel_spectrogram(mel_out.squeeze())
# wav = griffin_lim(magnitudes, stft.stft_fn, 60)
# save_wav(wav, os.path.join(args.output_dir, 'eval.wav'))
LOGGER.log(key="tacotron2_latency", value=measurements['tacotron2_time'])
LOGGER.log(key="latency", value=(measurements['tacotron2_time']))
LOGGER.iteration_stop()
LOGGER.finish()
def audio2mel2audio(dataset_path,
audiopaths_and_text,
melpaths_and_text,
args,
use_intermed=None):
melpaths_and_text_list = \
load_filepaths_and_text(dataset_path, melpaths_and_text)
audiopaths_and_text_list = \
load_filepaths_and_text(dataset_path, audiopaths_and_text)
# n = 10
# print(f"The first {n} melpaths and text are {melpaths_and_text_list[:n]}")
# print(f"The first {n} audiopaths and text are {audiopaths_and_text_list[:n]}")
# torchaudio implementation
spec = T.Spectrogram(
n_fft=args.filter_length,
win_length=args.win_length,
hop_length=args.hop_length,
power=1,
normalized=True,
)
# print(spec)
griffin_lim = T.GriffinLim(
n_fft=args.filter_length,
win_length=args.win_length,
hop_length=args.hop_length,
n_iter=args.n_iters,
power=1,
normalized=True,
)
# import pdb; pdb.set_trace()
print(args)
data_path = "/data/logotypografia_simple/cleaned_wavs/"
# tacotron-based implementations
# stft_fn = STFT(args.filter_length, args.hop_length, args.win_length)
# data_loader = TextMelLoader(dataset_path, audiopaths_and_text, args)
for i in range(len(melpaths_and_text_list)):
# specotrogram calculation based on internal components, buggy
# spec = data_loader.get_spec(audiopaths_and_text_list[i][0])
# wave = griffin_lim(spec, stft_fn, n_iters=30)
# wave = wave.detach().cpu().numpy()
# spectrogram calculation based on torchaudio
wav_name = data_path + audiopaths_and_text_list[i][0].split("/")[-1]
audio, sampling_rate = load_wav_to_torch(wav_name)
_spectrogram = spec(audio)
inv_waveform = griffin_lim(_spectrogram)
# torch.save(mel, f"grifin_lin/{}")
inv_wav_name = "griffin_lim_inv_audio_custom7/" \
+ audiopaths_and_text_list[i][0].split("/")[-1]
print(f"Saving reconstructed wav with name {inv_wav_name}")
write(inv_wav_name, 16000, inv_waveform.detach().cpu().numpy())