little-tts / scripts /render_training_samples.py
TuringLogic's picture
Little TTS initial release
b49ed64
Raw History Blame Contribute Delete
1.71 kB
"""Render consistent CPU listening previews from an inference checkpoint."""
import argparse
import json
from pathlib import Path
import soundfile as sf
import torch
from little_tts.infer import load_model, speak
from little_tts.text import set_frontend
TEXTS = (
"Hello! This is a sample of my voice as I learn to speak naturally.",
"The morning sunlight streamed through the window, warming the quiet room.",
)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--checkpoint", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
args = parser.parse_args()
torch.set_num_threads(2)
set_frontend("own")
model = load_model(args.checkpoint, "cpu")
args.output.mkdir(parents=True, exist_ok=True)
records = []
with torch.inference_mode():
for voice in model.config.speakers:
for index, text in enumerate(TEXTS, 1):
wave, phones, _ = speak(model, text, voice, seed=0)
peak = float(wave.abs().max())
# Attenuate peaks above .95 for playback, without boosting quiet output.
gain = min(1., .95 / max(peak, 1e-9))
path = args.output / f"voice_{voice}_{index}.wav"
sf.write(path, wave.numpy() * gain, model.config.sample_rate, subtype="PCM_16")
records.append(dict(voice=voice, text=text, file=path.name,
seconds=len(wave)/model.config.sample_rate, raw_peak=peak, playback_gain=gain))
(args.output / "samples.json").write_text(json.dumps(records, indent=2) + "\n")
print("Listening samples ready:", args.output, flush=True)
if __name__ == "__main__":
main()