NtTestAlert/tortoise-tts
1
0
Fork 0
forked from mrq/tortoise-tts
Code Issues Pull Requests Packages Projects Releases Wiki Activity

param improvements from investigation

Browse Source
This commit is contained in:
James Betker 2022-04-01 11:34:40 -06:00
parent 7f54f09a0d
commit 6721b85302
2 changed files with 11 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
eval_multiple.py
View File

@ -7,7 +7,7 @@ from utils.audio import load_audio
if __name__ == '__main__': if __name__ == '__main__':
fname = 'Y:\\libritts\\test-clean\\transcribed-brief-w2v.tsv' fname = 'Y:\\libritts\\test-clean\\transcribed-brief-w2v.tsv'
outpath = 'D:\\tmp\\tortoise-tts-eval\\redo_outlier' outpath = 'D:\\tmp\\tortoise-tts-eval\\attempt_best'
outpath_real = 'D:\\tmp\\tortoise-tts-eval\\real' outpath_real = 'D:\\tmp\\tortoise-tts-eval\\real'
os.makedirs(outpath, exist_ok=True) os.makedirs(outpath, exist_ok=True)
@ -24,8 +24,9 @@ if __name__ == '__main__':
path = os.path.join(os.path.dirname(fname), line[1]) path = os.path.join(os.path.dirname(fname), line[1])
cond_audio = load_audio(path, 22050) cond_audio = load_audio(path, 22050)
torchaudio.save(os.path.join(outpath_real, os.path.basename(line[1])), cond_audio, 22050) torchaudio.save(os.path.join(outpath_real, os.path.basename(line[1])), cond_audio, 22050)
sample = tts.tts(transcript, [cond_audio, cond_audio], num_autoregressive_samples=256, k=1, diffusion_iterations=200, cond_free=False, sample = tts.tts(transcript, [cond_audio, cond_audio], num_autoregressive_samples=512, k=1,
top_k=None, top_p=.95, typical_sampling=False, temperature=.7, length_penalty=.5, repetition_penalty=1) repetition_penalty=2.0, length_penalty=2, temperature=.5, top_p=.5,
diffusion_temperature=.7, cond_free_k=2, diffusion_iterations=400)
down = torchaudio.functional.resample(sample, 24000, 22050) down = torchaudio.functional.resample(sample, 24000, 22050)
fout_path = os.path.join(outpath, os.path.basename(line[1])) fout_path = os.path.join(outpath, os.path.basename(line[1]))
torchaudio.save(fout_path, down.squeeze(0), 22050) torchaudio.save(fout_path, down.squeeze(0), 22050)

13
sweep.py
View File

@ -25,18 +25,18 @@ def permutations(args):
if __name__ == '__main__': if __name__ == '__main__':
fname = 'Y:\\libritts\\test-clean\\transcribed-brief-w2v.tsv' fname = 'Y:\\libritts\\test-clean\\transcribed-brief-w2v.tsv'
outpath_base = 'D:\\tmp\\tortoise-tts-eval\\std_sweep_diffusion' outpath_base = 'D:\\tmp\\tortoise-tts-eval\\std_sweep3'
outpath_real = 'D:\\tmp\\tortoise-tts-eval\\real' outpath_real = 'D:\\tmp\\tortoise-tts-eval\\real'
arg_ranges = { arg_ranges = {
'diffusion_temperature': [.5, .7, 1], 'top_p': [.3,.4,.5,.6],
'cond_free_k': [.5, 1, 2], 'temperature': [.5, .6],
} }
cfgs = permutations(arg_ranges) cfgs = permutations(arg_ranges)
shuffle(cfgs) shuffle(cfgs)
for cfg in cfgs: for cfg in cfgs:
outpath = os.path.join(outpath_base, f'{cfg["cond_free_k"]}_{cfg["diffusion_temperature"]}') outpath = os.path.join(outpath_base, f'{cfg["top_p"]}_{cfg["temperature"]}')
os.makedirs(outpath, exist_ok=True) os.makedirs(outpath, exist_ok=True)
os.makedirs(outpath_real, exist_ok=True) os.makedirs(outpath_real, exist_ok=True)
with open(fname, 'r', encoding='utf-8') as f: with open(fname, 'r', encoding='utf-8') as f:
@ -51,8 +51,9 @@ if __name__ == '__main__':
path = os.path.join(os.path.dirname(fname), line[1]) path = os.path.join(os.path.dirname(fname), line[1])
cond_audio = load_audio(path, 22050) cond_audio = load_audio(path, 22050)
torchaudio.save(os.path.join(outpath_real, os.path.basename(line[1])), cond_audio, 22050) torchaudio.save(os.path.join(outpath_real, os.path.basename(line[1])), cond_audio, 22050)
sample = tts.tts(transcript, [cond_audio, cond_audio], num_autoregressive_samples=256, k=1, diffusion_iterations=200, cond_free=False, sample = tts.tts(transcript, [cond_audio, cond_audio], num_autoregressive_samples=256, k=1, diffusion_iterations=200,
repetition_penalty=1.5, length_penalty=2, temperature=.9, top_p=.9) repetition_penalty=2.0, length_penalty=2, temperature=.5, top_p=.5,
diffusion_temperature=.7, cond_free_k=2, **cfg)
down = torchaudio.functional.resample(sample, 24000, 22050) down = torchaudio.functional.resample(sample, 24000, 22050)
fout_path = os.path.join(outpath, os.path.basename(line[1])) fout_path = os.path.join(outpath, os.path.basename(line[1]))
torchaudio.save(fout_path, down.squeeze(0), 22050) torchaudio.save(fout_path, down.squeeze(0), 22050)