model.py

from transformers import pipeline
from accelerate import Accelerator
import spaces

model_id = "JacobLinCool/whisper-large-v3-turbo-common_voice_19_0-zh-TW"

pipe = None


def load_model():
    global pipe
    device = Accelerator().device
    pipe = pipeline("automatic-speech-recognition", model=model_id, device=device)


@spaces.GPU()
def transcribe_audio_local(audio: str) -> str:
    print(f"{audio=}")

    if pipe is None:
        load_model()

    out = pipe(audio, return_timestamps=True)
    print(f"{out=}")

    return out["text"]