app.py

import gradio as gr
from faster_whisper import WhisperModel

model_size = 'large-v3'

def load_model(model_size):
    if torch.cuda.is_available():
        model = WhisperModel(model_size, device="cuda", compute_type="float16")
        # model = WhisperModel(model_size, device="cuda", compute_type="int8_float16")
    else:
        model = WhisperModel(model_size, device="cpu", compute_type="int8")

def speech_to_text(audio_file, _model_size):
    global model_size, model
    if model_size != _model_size:
        model_size = _model_size
        model = load_model(model_size)

    with torch.no_grad():
        segments, info = model.transcribe(
            audio_file,
            verbose=True,
            language='japanese',
            beam_size=5,
            vad_filter=True,
            without_timestamps=False,
        )

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


load_model(model_size)

gr.Interface(
    fn=speech_to_text,
    inputs=[
        gr.Audio(sources="upload", type="filepath"),
        gr.Dropdown(value=model_size, choices=["tiny", "base", "small", "medium", "large", "large-v2", "large-v3"]),
        ],
    outputs="text").launch()