def build_phoneme_dicts(text_lang_pairs):
"""Create dictionaries (possibly more languages) of words (from a list of texts) with IPA equivalents."""
dictionaries = {}
Logger.progress(0 / len(text_lang_pairs),
prefix='Building phoneme dictionary:')
for i, (t, l) in enumerate(text_lang_pairs):
if not (l in dictionaries):
dictionaries[l] = {}
clear_words = remove_punctuation(t).split()
for w in clear_words:
if w in dictionaries[l]: continue
dictionaries[l][w] = _phonemize(w, l)[:-1]
Logger.progress((i + 1) / len(text_lang_pairs),
prefix='Building phoneme dictionary:')
return dictionaries
def create_meta_file(dataset_name, dataset_root_dir, output_metafile_name, audio_sample_rate, num_fft_freqs, spectrograms=True, phonemes=True):
"""Create the meta-file and spectrograms (mel and linear, optionally) or phonemized utterances (optionally).
Format details:
Every line of the metadata file contains info about one dataset item.
The line has following format
'id|speaker|language|audio_file_path|mel_spectrogram_path|linear_spectrogram_path|text|phonemized_text'
And the following must hold
'audio_file_path' can be empty if loading just spectrograms
'text' should be carefully normalized and should contain interpunction
'phonemized_text' can be empty if loading just raw text
Arguments:
dataset_name (string): Name of the dataset, loaders.py should contain a function for loading with a corresponding name.
dataset_root_dir (string): Root directory from which is the dataset build and to which are spectrograms and the meta-file saved..
output_metafile_name (string): Name of the output meta-file.
audio_sample_rate (int): Sample rate of audios, used if spectrograms is set True.
num_fft_freqs (int): Number of frequency bands used during spectrogram computation, used if spectrograms is set True.
Keyword arguments:
spectrograms (boolean, default True): If true, spetrograms (both mel and linear) are computed and saved.
phonemes (boolean, default True): If true, phonemized variants of utterances are computed and saved.
"""
# save current sample rate and fft freqs hyperparameters, as we may process dataset with different sample rate
if spectrograms:
old_sample_rate = hp.sample_rate
hp.sample_rate = audio_sample_rate
old_fft_freqs = hp.num_fft
hp.num_fft = num_fft_freqs
# load metafiles, an item is a list like: [text, audiopath, speaker_id, language_code]
items = loaders.get_loader_by_name(dataset_name)(dataset_root_dir)
# build dictionaries for translation to IPA from source languages, see utils.text for details
if phonemes:
text_lang_pairs = [(i[0], hp.languages[0] if i[3] == "" else i[3]) for i in items]
phoneme_dicts = text.build_phoneme_dicts(text_lang_pairs)
# prepare directories which will store spectrograms
if spectrograms:
spectrogram_dirs = [os.path.join(dataset_root_dir, 'spectrograms'),
os.path.join(dataset_root_dir, 'linear_spectrograms')]
for x in spectrogram_dirs:
if not os.path.exists(x): os.makedirs(x)
# iterate through items and build the meta-file
metafile_path = os.path.join(dataset_root_dir, output_metafile_name)
with open(metafile_path, 'w', encoding='utf-8') as f:
Logger.progress(0, prefix='Building metafile:')
for i in range(len(items)):
raw_text, audio_path, speaker, language = items[i]
if language == "": language = hp.languages[0]
phonemized_text = text.to_phoneme(raw_text, False, language, phoneme_dicts[language]) if phonemes else ""
spectrogram_paths = "|"
if spectrograms:
spec_name = f'{str(i).zfill(6)}.npy'
audio_data = audio.load(os.path.join(dataset_root_dir, audio_path))
np.save(os.path.join(spectrogram_dirs[0], spec_name), audio.spectrogram(audio_data, True))
np.save(os.path.join(spectrogram_dirs[1], spec_name), audio.spectrogram(audio_data, False))
spectrogram_paths = os.path.join('spectrograms', spec_name) + '|' + os.path.join('linear_spectrograms', spec_name)
print(f'{str(i).zfill(6)}|{speaker}|{language}|{audio_path}|{spectrogram_paths}|{raw_text}|{phonemized_text}', file=f)
Logger.progress((i + 1) / len(items), prefix='Building metafile:')
# restore the original sample rate and fft freq values
if spectrograms:
hp.sample_rate = old_sample_rate
hp.num_fft = old_fft_freqs
