2022-03-29 01:33:31 +00:00
|
|
|
import os
|
|
|
|
|
|
|
|
import torchaudio
|
|
|
|
|
|
|
|
from api import TextToSpeech
|
|
|
|
from utils.audio import load_audio
|
|
|
|
|
|
|
|
if __name__ == '__main__':
|
|
|
|
fname = 'Y:\\libritts\\test-clean\\transcribed-brief-w2v.tsv'
|
2022-04-13 02:53:09 +00:00
|
|
|
outpath = 'D:\\tmp\\tortoise-tts-eval\\diverse_new_decoder_1'
|
2022-03-29 01:33:31 +00:00
|
|
|
outpath_real = 'D:\\tmp\\tortoise-tts-eval\\real'
|
|
|
|
|
|
|
|
os.makedirs(outpath, exist_ok=True)
|
|
|
|
os.makedirs(outpath_real, exist_ok=True)
|
|
|
|
with open(fname, 'r', encoding='utf-8') as f:
|
|
|
|
lines = [l.strip().split('\t') for l in f.readlines()]
|
|
|
|
|
|
|
|
recorder = open(os.path.join(outpath, 'transcript.tsv'), 'w', encoding='utf-8')
|
|
|
|
tts = TextToSpeech()
|
|
|
|
for e, line in enumerate(lines):
|
|
|
|
transcript = line[0]
|
|
|
|
if len(transcript) > 120:
|
|
|
|
continue # We need to support this, but cannot yet.
|
|
|
|
path = os.path.join(os.path.dirname(fname), line[1])
|
|
|
|
cond_audio = load_audio(path, 22050)
|
|
|
|
torchaudio.save(os.path.join(outpath_real, os.path.basename(line[1])), cond_audio, 22050)
|
2022-04-12 22:40:42 +00:00
|
|
|
sample = tts.tts(transcript, [cond_audio, cond_audio], num_autoregressive_samples=256, k=1,
|
2022-04-01 17:34:40 +00:00
|
|
|
repetition_penalty=2.0, length_penalty=2, temperature=.5, top_p=.5,
|
2022-04-12 22:40:42 +00:00
|
|
|
diffusion_temperature=.7, cond_free_k=2, diffusion_iterations=100)
|
2022-04-10 20:41:13 +00:00
|
|
|
|
2022-03-29 01:33:31 +00:00
|
|
|
down = torchaudio.functional.resample(sample, 24000, 22050)
|
2022-04-11 05:19:15 +00:00
|
|
|
fout_path = os.path.join(outpath, os.path.basename(line[1]))
|
2022-03-29 01:33:31 +00:00
|
|
|
torchaudio.save(fout_path, down.squeeze(0), 22050)
|
2022-04-10 20:41:13 +00:00
|
|
|
|
2022-03-29 01:33:31 +00:00
|
|
|
recorder.write(f'{transcript}\t{fout_path}\n')
|
|
|
|
recorder.flush()
|
|
|
|
recorder.close()
|