File size: 6,034 Bytes
8a2aca3 08238e0 e154110 9b9c715 8a2aca3 9b9c715 8a2aca3 9b9c715 fda2aa0 8a2aca3 9b9c715 8a2aca3 9b9c715 8a2aca3 9b9c715 08238e0 8a2aca3 9b9c715 08238e0 9b9c715 08238e0 e154110 08238e0 e154110 08238e0 e154110 08238e0 e154110 08238e0 e154110 08238e0 e154110 08238e0 e154110 08238e0 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 |
# Synthesize all Harvard Lists 77x lists of 10x sentences to single .wav
# 1. using mimic3 english 1x/4x non-english 1x/4x
# Call visualize_tts_plesantness.py for 4figs [eng 1x/4x vs human, non-eng 1x/4x vs human-libri]
import soundfile
import json
import numpy as np
import audb
from pathlib import Path
LABELS = ['arousal', 'dominance', 'valence']
def load_speech(split=None):
DB = [
# [dataset, version, table, has_timdeltas_or_is_full_wavfile]
# ['crema-d', '1.1.1', 'emotion.voice.test', False],
['librispeech', '3.1.0', 'train-clean-360', False],
# ['emodb', '1.2.0', 'emotion.categories.train.gold_standard', False],
# ['entertain-playtestcloud', '1.1.0', 'emotion.categories.train.gold_standard', True],
# ['erik', '2.2.0', 'emotion.categories.train.gold_standard', True],
# ['meld', '1.3.1', 'emotion.categories.train.gold_standard', False],
# ['msppodcast', '5.0.0', 'emotion.categories.train.gold_standard', False], # tandalone bucket because it has gt labels?
# ['myai', '1.0.1', 'emotion.categories.train.gold_standard', False],
# ['casia', None, 'emotion.categories.gold_standard', False],
# ['switchboard-1', None, 'sentiment', True],
# ['swiss-parliament', None, 'segments', True],
# ['argentinian-parliament', None, 'segments', True],
# ['austrian-parliament', None, 'segments', True],
# #'german', --> bundestag
# ['brazilian-parliament', None, 'segments', True],
# ['mexican-parliament', None, 'segments', True],
# ['portuguese-parliament', None, 'segments', True],
# ['spanish-parliament', None, 'segments', True],
# ['chinese-vocal-emotions-liu-pell', None, 'emotion.categories.desired', False],
# peoples-speech slow
# ['peoples-speech', None, 'train-initial', False]
]
output_list = []
for database_name, ver, table, has_timedeltas in DB:
a = audb.load(database_name,
sampling_rate=16000,
format='wav',
mixdown=True,
version=ver,
cache_root='/cache/audb/')
a = a[table].get()
if has_timedeltas:
print(f'{has_timedeltas=}')
# a = a.reset_index()[['file', 'start', 'end']]
# output_list += [[*t] for t
# in zip(a.file.values, a.start.dt.total_seconds().values, a.end.dt.total_seconds().values)]
else:
output_list += [f for f in a.index] # use file (no timedeltas)
return output_list
natural_wav_paths = load_speech()
# SYNTHESIZE mimic mimicx4 crema-d
import msinference
import os
with open('harvard.json', 'r') as f:
harvard_individual_sentences = json.load(f)['sentences']
synthetic_wav_paths = ['./style_vector/' + i for i in
os.listdir('./style_vector/')]
synthetic_wav_paths_4x = ['./style_vector_v2/' + i for i in
os.listdir('./style_vector_v2/')]
synthetic_wav_paths_foreign = ['./mimic3_foreign/' + i for i in os.listdir('./mimic3_foreign/')]
synthetic_wav_paths_foreign_4x = ['./mimic3_foreign_4x/' + i for i in os.listdir('./mimic3_foreign_4x/')]
for audio_prompt in ['english',
'english_4x',
'human',
'foreign',
'foreign_4x']:
if not os.path.isfile(f'{audio_prompt}_z.wav'):
total_audio = []
ix = 0
for list_of_10 in harvard_individual_sentences[:10000]:
# long_sentence = ' '.join(list_of_10['sentences'])
# harvard.append(long_sentence.replace('.', ' '))
for text in list_of_10['sentences']:
if audio_prompt == 'english':
style_vec = msinference.compute_style(
synthetic_wav_paths[ix % 134])
elif audio_prompt == 'english_4x':
style_vec = msinference.compute_style(
synthetic_wav_paths_4x[ix % 134])
elif audio_prompt == 'human':
style_vec = msinference.compute_style(
natural_wav_paths[ix % len(natural_wav_paths)])
elif audio_prompt == 'foreign':
style_vec = msinference.compute_style(
synthetic_wav_paths_foreign[ix % 204])
elif audio_prompt == 'foreign_4x':
style_vec = msinference.compute_style(
synthetic_wav_paths_foreign_4x[ix % 204])
else:
print('unknonw list of style vecto')
print(ix, text)
ix += 1
x = msinference.inference(text,
style_vec,
alpha=0.3,
beta=0.7,
diffusion_steps=7,
embedding_scale=1)
total_audio.append(x)
# concat before write
# -- for 10x sentenctes
print('_____________________')
# -- for 77x lists
total_audio = np.concatenate(total_audio)
soundfile.write(f'{audio_prompt}_z.wav', total_audio, 24000)
else:
print('\nALREADY EXISTS\n') |