fn.py

from faster_whisper import WhisperModel

model = None
model_size = None

def load_model(_model_size):
    global model_size, model

    if _model_size and model_size != _model_size:
        model_size = _model_size

    try:
        model = WhisperModel(model_size, device="cuda", compute_type="float16")
    except:
        model = WhisperModel(model_size, device="cpu", compute_type="int8")

def speech_to_text(audio_file, _model_size = None):
    global model_size, model

    load_model(_model_size)

    segments, info = model.transcribe(
        audio_file,
        language='ja',
        beam_size=5,
        vad_filter=True,
        without_timestamps=False,
    )

    text_only = ''
    text_with_timestamps = ''
    for segment in segments:
        text_only += f"{segment.text}\n"
        text_with_timestamps += f"{segment.start:.2f}\t{segment.end:.2f}\t{segment.text}\n"

    return text_only, text_with_timestamps