tiny-voca-hindi / speak.py
RL10x's picture
Add inference code (Python + C++ engine + web page)
ab1b068 verified
Raw History Blame Contribute Delete
2.71 kB
"""Hindi text -> WAV with the PyTorch checkpoint.
python speak.py "भारत एक विशाल देश है, जहाँ अनेक भाषाएँ बोली जाती हैं।" --out out.wav
python speak.py "आज 15 अगस्त है और मौसम 32 डिग्री है।" --style happy --out happy.wav
python speak.py --text-file story.txt --style conv --speed 1.1 --out story.wav
Text is normalised first (numbers, dates, money, units and common English words -> Hindi words), then read sentence by
sentence. Styles: neutral, conv, happy, sad, anger, fear, surprise, disgust.
"""
import argparse
import sys
import time
from pathlib import Path
import soundfile as sf
sys.path.insert(0, str(Path(__file__).resolve().parent))
from hindi_text import normalize_hi # noqa: E402
from tiny_voca.infer import limit_peak, load_model, speak_text # noqa: E402
STYLES = ("neutral", "conv", "happy", "sad", "anger", "fear", "surprise", "disgust")
def main():
ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("text", nargs="?", help="Hindi text (Devanagari)")
ap.add_argument("--text-file", type=Path)
ap.add_argument("--style", choices=STYLES, default="neutral")
ap.add_argument("--speed", type=float, default=1.0, help="0.5 (slow) to 2 (fast)")
here = Path(__file__).resolve().parent
default = next((p for p in (here / "models" / "tiny_voca_hindi.pt", here / "tiny_voca_hindi.pt") if p.exists()),
here / "models" / "tiny_voca_hindi.pt")
ap.add_argument("--model", type=Path, default=default)
ap.add_argument("--out", type=Path, default=Path("out.wav"))
ap.add_argument("--device", default="cpu")
a = ap.parse_args()
text = a.text if a.text is not None else (a.text_file.read_text(encoding="utf-8") if a.text_file else None)
if not text:
ap.error("give the text or --text-file")
if not a.model.exists():
sys.exit(f"{a.model} not found: run python download_models.py first")
model = load_model(a.model, a.device)
t0 = time.time()
spoken = normalize_hi(text, english=True).replace("।", ".")
wave, _ = speak_text(model, spoken, a.style, speed=a.speed, peak=None)
wave = limit_peak(wave)
seconds = len(wave) / model.config.sample_rate
a.out.parent.mkdir(parents=True, exist_ok=True)
sf.write(a.out, wave.numpy(), model.config.sample_rate, subtype="PCM_16")
took = time.time() - t0
print(f"{a.out}: {seconds:.2f} s of audio in {took:.2f} s ({seconds / took:.0f}x real time)")
if __name__ == "__main__":
main()