inference.py

"""Minimal transcription example for GigaAM-He.

    pip install torch torchaudio soundfile numpy
    pip install git+https://github.com/salute-developers/GigaAM.git
    python inference.py clip.wav

For anything longer than ~20 s use server.py, which segments first: the model
is trained on segments up to 20 s and a longer input degrades rather than
failing loudly.
"""
import sys

import gigaam

CHECKPOINT = "gigaam-he-twostage.ckpt"


def main() -> None:
    if len(sys.argv) < 2:
        print(__doc__)
        raise SystemExit(2)
    model = gigaam.load_model(CHECKPOINT, device="cuda", fp16_encoder=False)
    for path in sys.argv[1:]:
        result = model.transcribe(path)
        text = getattr(result, "transcription", None) or getattr(result, "text", str(result))
        print(f"{path}: {text}")


if __name__ == "__main__":
    main()