def audio2mel(audiopaths_and_text, melpaths_and_text, args):
melpaths_and_text_list = load_filepaths_and_text(melpaths_and_text)
audiopaths_and_text_list = load_filepaths_and_text(audiopaths_and_text)
data_loader = TextMelLoader(audiopaths_and_text, args)
for i in range(len(melpaths_and_text_list)):
if i % 100 == 0:
print("done", i, "/", len(melpaths_and_text_list))
mel = data_loader.get_mel(audiopaths_and_text_list[i][0])
torch.save(mel, melpaths_and_text_list[i][0])
def __init__(self,
dataset_path,
audiopaths_and_text,
text_cleaners,
n_mel_channels,
symbol_set='english_basic',
n_speakers=1,
load_mel_from_disk=True,
max_wav_value=None,
sampling_rate=None,
filter_length=None,
hop_length=None,
win_length=None,
mel_fmin=None,
mel_fmax=None,
**ignored):
self.audiopaths_and_text = load_filepaths_and_text(
dataset_path, audiopaths_and_text,
has_speakers=(n_speakers > 1))
self.load_mel_from_disk = load_mel_from_disk
if not load_mel_from_disk:
self.max_wav_value = max_wav_value
self.sampling_rate = sampling_rate
self.stft = layers.TacotronSTFT(
filter_length, hop_length, win_length,
n_mel_channels, sampling_rate, mel_fmin, mel_fmax)
def __init__(self, dataset_path, audiopaths_and_text, args):
self.audiopaths_and_text = load_filepaths_and_text(dataset_path, audiopaths_and_text)
self.max_wav_value = args.max_wav_value
self.sampling_rate = args.sampling_rate
self.stft = layers.TacotronSTFT(
args.filter_length, args.hop_length, args.win_length,
args.n_mel_channels, args.sampling_rate, args.mel_fmin,
args.mel_fmax)
self.segment_length = args.segment_length
random.seed(1234)
random.shuffle(self.audiopaths_and_text)
def __init__(self, dataset_path, audiopaths_and_text, args, load_mel_from_disk=True):
self.audiopaths_and_text = load_filepaths_and_text(dataset_path, audiopaths_and_text)
self.text_cleaners = args.text_cleaners
self.load_mel_from_disk = load_mel_from_disk
if not load_mel_from_disk:
self.max_wav_value = args.max_wav_value
self.sampling_rate = args.sampling_rate
self.stft = layers.TacotronSTFT(
args.filter_length, args.hop_length, args.win_length,
args.n_mel_channels, args.sampling_rate, args.mel_fmin,
args.mel_fmax)
def __init__(self, audiopaths_and_text, args):
self.audiopaths_and_text = load_filepaths_and_text(audiopaths_and_text)
self.text_cleaners = args.text_cleaners
self.max_wav_value = args.max_wav_value
self.sampling_rate = args.sampling_rate
self.load_mel_from_disk = args.load_mel_from_disk
self.stft = layers.TacotronSTFT(args.filter_length, args.hop_length,
args.win_length, args.n_mel_channels,
args.sampling_rate, args.mel_fmin,
args.mel_fmax)
random.seed(1234)
random.shuffle(self.audiopaths_and_text)
def audio2mel(dataset_path: str, audiopaths_and_text: str,
melpaths_and_text: str, args: ArgumentParser) -> None:
"""Create mel spectrograms on disk from audio files.
Args:
dataset_path (str): Path to dataset
audiopaths_and_text (str): Path to filelist with audio paths and text
melpaths_and_text (str): Path to filelist with mel paths and text
args (ArgumentParser): Namespace with arguments
"""
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)
data_loader = TextMelLoader(dataset_path, audiopaths_and_text, args)
for i, melpath_and_text in enumerate(melpaths_and_text_list):
if i % 100 == 0:
print("done", i, "/", len(melpaths_and_text_list))
mel = data_loader.get_mel(audiopaths_and_text_list[i][0])
torch.save(mel, melpath_and_text[0])
def audio2mel(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]}")
data_loader = TextMelLoader(dataset_path, audiopaths_and_text, args)
for i in range(len(melpaths_and_text_list)):
if i % 100 == 0:
print("done", i, "/", len(melpaths_and_text_list))
mel = data_loader.get_mel(audiopaths_and_text_list[i][0])
torch.save(mel, melpaths_and_text_list[i][0])
def __init__(self, dataset_path, audiopaths_and_text, segment_length,
n_mel_channels, max_wav_value, sampling_rate, filter_length,
hop_length, win_length, mel_fmin, mel_fmax, args):
self.audiopaths_and_text = load_filepaths_and_text(
dataset_path, audiopaths_and_text)
self.max_wav_value = max_wav_value
self.sampling_rate = sampling_rate
self.stft = layers.TacotronSTFT(filter_length, hop_length, win_length,
n_mel_channels, sampling_rate,
mel_fmin, mel_fmax)
self.segment_length = segment_length
random.seed(1234)
random.shuffle(self.audiopaths_and_text)
def __init__(self, dataset_path, audiopaths_and_text, args, speaker_ids=None):
self.audiopaths_and_text = load_filepaths_and_text(dataset_path, audiopaths_and_text)
self.text_cleaners = args.text_cleaners
self.max_wav_value = args.max_wav_value
self.sampling_rate = args.sampling_rate
self.load_mel_from_disk = args.load_mel_from_disk
self.stft = layers.TacotronSTFT(
args.filter_length, args.hop_length, args.win_length,
args.n_mel_channels, args.sampling_rate, args.mel_fmin,
args.mel_fmax)
random.seed(1234)
random.shuffle(self.audiopaths_and_text)
self.speaker_ids = speaker_ids
if speaker_ids is None:
self.speaker_ids = self.create_speaker_lookup_table(self.audiopaths_and_text)
def __init__(
self,
dataset_path,
audiopaths_and_text,
text_cleaners,
n_mel_channels,
symbol_set='english_basic',
p_arpabet=1.0,
n_speakers=1,
load_mel_from_disk=True,
load_pitch_from_disk=True,
pitch_mean=214.72203, # LJSpeech defaults
pitch_std=65.72038,
max_wav_value=None,
sampling_rate=None,
filter_length=None,
hop_length=None,
win_length=None,
mel_fmin=None,
mel_fmax=None,
prepend_space_to_text=False,
append_space_to_text=False,
pitch_online_dir=None,
betabinomial_online_dir=None,
use_betabinomial_interpolator=True,
pitch_online_method='pyin',
**ignored):
# Expect a list of filenames
if type(audiopaths_and_text) is str:
audiopaths_and_text = [audiopaths_and_text]
self.dataset_path = dataset_path
self.audiopaths_and_text = load_filepaths_and_text(
dataset_path, audiopaths_and_text, has_speakers=(n_speakers > 1))
self.load_mel_from_disk = load_mel_from_disk
if not load_mel_from_disk:
self.max_wav_value = max_wav_value
self.sampling_rate = sampling_rate
self.stft = layers.TacotronSTFT(filter_length, hop_length,
win_length, n_mel_channels,
sampling_rate, mel_fmin, mel_fmax)
self.load_pitch_from_disk = load_pitch_from_disk
self.prepend_space_to_text = prepend_space_to_text
self.append_space_to_text = append_space_to_text
assert p_arpabet == 0.0 or p_arpabet == 1.0, (
'Only 0.0 and 1.0 p_arpabet is currently supported. '
'Variable probability breaks caching of betabinomial matrices.')
self.tp = TextProcessing(symbol_set,
text_cleaners,
p_arpabet=p_arpabet)
self.n_speakers = n_speakers
self.pitch_tmp_dir = pitch_online_dir
self.f0_method = pitch_online_method
self.betabinomial_tmp_dir = betabinomial_online_dir
self.use_betabinomial_interpolator = use_betabinomial_interpolator
if use_betabinomial_interpolator:
self.betabinomial_interpolator = BetaBinomialInterpolator()
expected_columns = (2 + int(load_pitch_from_disk) + (n_speakers > 1))
assert not (load_pitch_from_disk and self.pitch_tmp_dir is not None)
if len(self.audiopaths_and_text[0]) < expected_columns:
raise ValueError(
f'Expected {expected_columns} columns in audiopaths file. '
'The format is <mel_or_wav>|[<pitch>|]<text>[|<speaker_id>]')
if len(self.audiopaths_and_text[0]) > expected_columns:
print('WARNING: Audiopaths file has more columns than expected')
to_tensor = lambda x: torch.Tensor([x]) if type(x) is float else x
self.pitch_mean = to_tensor(pitch_mean)
self.pitch_std = to_tensor(pitch_std)
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())
