app.py

import torch
import gradio as gr
from pytube import YouTube
from transformers import pipeline

MODEL_NAME = "yuweiiizz/whisper-small-taiwanese"
lang = "chinese"

# 根據是否有可用的 CUDA 設備來選擇設備
device = 0 if torch.cuda.is_available() else "cpu"

# 初始化 pipeline，指定任務、模型和設備
pipe = pipeline(
    task="automatic-speech-recognition",
    chunk_length_s=15,
    model=MODEL_NAME,
    device=device,
)

# 設置模型的語言和任務
pipe.model.config.forced_decoder_ids = pipe.tokenizer.get_decoder_prompt_ids(language=lang, task="transcribe")

# 定義轉錄功能
def transcribe(microphone=None, file_upload=None):
    warn_output = ""
    if microphone is not None and file_upload is not None:
        warn_output = "警告：您同時使用了麥克風與上傳音訊檔案，將只會使用麥克風錄製的檔案。\n"
    elif microphone is None and file_upload is None:
        return "錯誤：您必須至少使用麥克風或上傳一個音頻檔案。"

    file = microphone if microphone is not None else file_upload
    text = pipe(file)["text"]
    return warn_output + text

# 定義 YouTube 轉寫功能
def yt_transcribe(yt_url):
    yt = YouTube(yt_url)
    stream = yt.streams.filter(only_audio=True).first()
    stream.download(filename="audio.mp3")
    text = pipe("audio.mp3")["text"]
    # 嵌入 YouTube 影片
    video_id = yt_url.split("?v=")[-1]
    html_embed = f'<center><iframe width="500" height="320" src="https://www.youtube.com/embed/{video_id}"></iframe></center>'
    
    return html_embed, text

# 初始化 Gradio Blocks
demo = gr.Blocks()

# 定義兩個介面
mf_transcribe = gr.Interface(
    fn=transcribe,
    inputs=gr.Audio(label="audio",type="filepath"),
    outputs="text",
    title="Whisper 演示: 語音轉錄",
    description=f"演示使用 fine-tuned checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME} 以及 🤗 Transformers 轉錄任意長度的音訊檔案",
    allow_flagging="manual",
)

yt_transcribe = gr.Interface(
    fn=yt_transcribe,
    inputs=[gr.Textbox(lines=1, placeholder="在此處貼上 YouTube 影片的 URL", label="YouTube URL")],
    outputs=["html", "text"],
    title="Whisper 演示: Youtube轉錄",
    description=f"演示使用 fine-tuned checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME} 以及 🤗 Transformers 轉錄任意長度的Youtube影片",
    allow_flagging="manual",
)

# 將兩個介面加入到標籤介面中
with demo:
    gr.TabbedInterface([mf_transcribe, yt_transcribe], ["語音轉錄", "Youtube轉錄"])

# 啟動並分享 Gradio 介面
demo.launch(share=True)