def get_final_ds_from_data(ds_data: DsDataList, text_data: TextDataList,
wav_data: WavDataList, mel_data: MelDataList,
wav_dir: Path, mel_dir: Path) -> FinalDsEntryList:
res = WavDataList()
for ds_data_entry, text_data_entry, wav_data_entry, mel_data_entry in zip(
ds_data.items(), text_data.items(), wav_data.items(),
mel_data.items()):
assert ds_data_entry.entry_id == text_data_entry.entry_id == wav_data_entry.entry_id == mel_data_entry.entry_id
assert ds_data_entry.symbols_language == text_data_entry.symbols_language
new_entry = FinalDsEntry(
entry_id=ds_data_entry.entry_id,
basename=ds_data_entry.basename,
speaker_gender=ds_data_entry.speaker_gender,
speaker_name=ds_data_entry.speaker_name,
symbols_language=ds_data_entry.symbols_language,
symbols_original=ds_data_entry.symbols,
symbols_original_format=ds_data_entry.symbols_format,
symbols=text_data_entry.symbols,
symbols_format=text_data_entry.symbols_format,
wav_original_absolute_path=ds_data_entry.wav_absolute_path,
wav_absolute_path=wav_dir / wav_data_entry.wav_relative_path,
wav_duration=wav_data_entry.wav_duration,
wav_sampling_rate=wav_data_entry.wav_sampling_rate,
mel_absolute_path=mel_dir / mel_data_entry.mel_relative_path,
mel_n_channels=mel_data_entry.mel_n_channels,
)
res.append(new_entry)
return res
def _save_speaker_examples(ds_dir: Path, examples: DsDataList,
logger: Logger) -> None:
logger.info("Saving examples for each speaker...")
example_dir = get_ds_examples_dir(ds_dir)
example_dir.mkdir(exist_ok=True, parents=True)
for i, example in enumerate(examples.items(True), start=1):
dest_file_name = f"{i}-{str(example.speaker_gender)}-{convert_to_ascii(example.speaker_name)}.wav"
dest_path = example_dir / dest_file_name
copyfile(example.wav_absolute_path, dest_path)
def preprocess(data: DsDataList, dest_dir: Path, n_jobs: int) -> WavDataList:
assert dest_dir.is_dir()
mt_method = partial(
preprocess_entry,
dest_dir=dest_dir,
entries_count=len(data),
)
with ThreadPoolExecutor(max_workers=n_jobs) as ex:
result = WavDataList(tqdm(ex.map(mt_method, data.items()), total=len(data)))
return result
def log_stats(ds_data: DsDataList, text_data: TextDataList):
stats: List[str, int, float, float, float] = []
text_lengths = [len(x.symbols) for x in text_data.items()]
stats.append((
"Overall",
len(text_lengths),
min(text_lengths),
max(text_lengths),
mean(text_lengths),
sum(text_lengths),
))
speakers_text_lengths: Dict[Speaker, List[float]] = {}
for ds_entry, text_entry in zip(ds_data.items(), text_data.items()):
if ds_entry.speaker_name not in speakers_text_lengths:
speakers_text_lengths[ds_entry.speaker_name] = []
speakers_text_lengths[ds_entry.speaker_name].append(
len(text_entry.symbols))
for speaker, speaker_text_lengths in speakers_text_lengths.items():
stats.append((
speaker,
len(speaker_text_lengths),
min(speaker_text_lengths),
max(speaker_text_lengths),
mean(speaker_text_lengths),
sum(speaker_text_lengths),
))
stats.sort(key=lambda x: (x[-1]), reverse=True)
stats_csv = pd.DataFrame(stats,
columns=[
"Speaker",
"# Entries",
"# Min",
"# Max",
"# Avg",
"# Total",
])
logger = getLogger(__name__)
with pd.option_context(
'display.max_rows',
None,
'display.max_columns',
None,
'display.width',
None,
'display.precision',
0,
):
logger.info(stats_csv)
def preprocess(data: DsDataList) -> TextDataList:
result = TextDataList()
for entry in data.items(True):
text_entry = TextData(
entry_id=entry.entry_id,
symbols=entry.symbols,
symbols_format=entry.symbols_format,
symbols_language=entry.symbols_language,
)
result.append(text_entry)
return result
def log_stats(ds_data: DsDataList, wav_data: WavDataList):
logger = getLogger(__name__)
if len(wav_data) > 0:
logger.info(f"Sampling rate: {wav_data.items()[0].wav_sampling_rate}")
stats: List[str, int, float, float, float, int] = []
durations = [x.wav_duration for x in wav_data.items()]
stats.append((
"Overall",
len(wav_data),
min(durations),
max(durations),
mean(durations),
sum(durations) / 60,
sum(durations) / 3600,
))
speaker_durations: Dict[Speaker, List[float]] = {}
for ds_entry, wav_entry in zip(ds_data.items(), wav_data.items()):
if ds_entry.speaker_name not in speaker_durations:
speaker_durations[ds_entry.speaker_name] = []
speaker_durations[ds_entry.speaker_name].append(wav_entry.wav_duration)
for speaker_name, speaker_durations in speaker_durations.items():
stats.append((
speaker_name,
len(speaker_durations),
min(speaker_durations),
max(speaker_durations),
mean(speaker_durations),
sum(speaker_durations) / 60,
sum(speaker_durations) / 3600,
))
stats.sort(key=lambda x: (x[-2]), reverse=True)
stats_csv = pd.DataFrame(stats, columns=[
"Speaker",
"# Entries",
"Min (s)",
"Max (s)",
"Avg (s)",
"Total (min)",
"Total (h)",
])
with pd.option_context(
'display.max_rows', None,
'display.max_columns', None,
'display.width', None,
'display.precision', 4,
):
print(stats_csv)
def process(data: WavDataList, ds: DsDataList, wav_dir: Path,
custom_hparams: Optional[Dict[str, str]],
save_callback: Callable[[WavData, DsData], Path]) -> List[Path]:
hparams = TSTFTHParams()
hparams = overwrite_custom_hparams(hparams, custom_hparams)
mel_parser = TacotronSTFT(hparams, logger=getLogger())
all_paths: List[Path] = []
for wav_entry, ds_entry in zip(data.items(True), ds.items(True)):
absolute_wav_path = wav_dir / wav_entry.wav_relative_path
mel_tensor = mel_parser.get_mel_tensor_from_file(absolute_wav_path)
absolute_path = save_callback(wav_entry=wav_entry,
ds_entry=ds_entry,
mel_tensor=mel_tensor)
all_paths.append(absolute_path)
return all_paths