app.py

# import os
# import time
# import json
# import gradio as gr
# import torch
# import torchaudio
# import numpy as np
# from denoiser.demucs import Demucs
# from pydub import AudioSegment

# modelpath = './denoiser/master64.th'

# def transcribe(file_upload, microphone):
#     file = microphone if microphone is not None else file_upload
#     model = Demucs(hidden=64)
#     state_dict = torch.load(modelpath, map_location='cpu')
#     model.load_state_dict(state_dict)
#     demucs = model
#     x, sr = torchaudio.load(file)
#     out = demucs(x[None])[0]
#     out = out / max(out.abs().max().item(), 1)
#     torchaudio.save('enhanced.wav', out, sr)
#     enhanced = AudioSegment.from_wav('enhanced.wav') # 只有去完噪的需要降 bitrate 再做語音識別
#     enhanced.export('enhanced.wav', format="wav", bitrate="256k")
#     return "enhanced.wav"

import os
import time
import json
import gradio as gr
import torch
import torchaudio
import numpy as np
from denoiser.demucs import Demucs
from pydub import AudioSegment
import soundfile as sf
import librosa

modelpath = './denoiser/master64.th'

def transcribe(file_upload, microphone):
    file = microphone if microphone is not None else file_upload

    # 新增音訊預處理 → 統一格式
    def preprocess_audio(path):
        data, sr = sf.read(path)
        
        # 如果是雙聲道 → 轉單聲道
        if len(data.shape) > 1:
            data = data.mean(axis=1)

        # 如果不是 16kHz → 重採樣
        if sr != 16000:
            data = librosa.resample(data, orig_sr=sr, target_sr=16000)
            sr = 16000

        # 儲存為 WAV 供模型使用
        sf.write("enhanced.wav", data, sr)
        return "enhanced.wav"

    # 如果是 MP3，先轉成 WAV 再處理
    if file.lower().endswith(".mp3"):
        audio = AudioSegment.from_file(file)
        audio = audio.set_frame_rate(16000).set_channels(1)  # 轉單聲道 + 16kHz
        audio.export("enhanced.wav", format="wav")
        file = "enhanced.wav"
    else:
        file = preprocess_audio(file)

    model = Demucs(hidden=64)
    state_dict = torch.load(modelpath, map_location='cpu')
    model.load_state_dict(state_dict)
    demucs = model.eval()

    x, sr = torchaudio.load(file)
    x = x[0:1]  # 強制取第一個聲道（確保是單聲道）

    with torch.no_grad():
        out = demucs(x[None])[0]

    out = out / max(out.abs().max().item(), 1)
    torchaudio.save('enhanced_final.wav', out, sr)

    # 輸出 WAV 格式給前端播放
    enhanced = AudioSegment.from_wav('enhanced_final.wav')
    enhanced.export('enhanced_final.mp3', format="mp3", bitrate="256k")

    return "enhanced_final.mp3"  # 回傳 MP3 更省空間
    
demo = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.Audio(type="filepath", label="語音質檢原始音檔"),
    ],
    outputs=gr.Audio(type="filepath", label="Output"),
    title="<h1>語音質檢/噪音去除 (語音增強)</h1>",
    description="""<h2><a href='https://www.twman.org' target='_blank'>TonTon Huang Ph.D.</a> | <a href='https://blog.twman.org/p/deeplearning101.html' target='_blank'>手把手帶你一起踩AI坑</a><br></h2><br>
                為了提升語音識別的效果，可以在識別前先進行噪音去除<br>
                    <a href='https://github.com/Deep-Learning-101' target='_blank'>Deep Learning 101 Github</a> | <a href='http://deeplearning101.twman.org' target='_blank'>Deep Learning 101</a> | <a href='https://www.facebook.com/groups/525579498272187/' target='_blank'>台灣人工智慧社團 FB</a> | <a href='https://www.youtube.com/c/DeepLearning101' target='_blank'>YouTube</a><br>
                    <a href='https://blog.twman.org/2025/03/AIAgent.html' target='_blank'>那些 AI Agent 要踩的坑</a>：探討多種 AI 代理人工具的應用經驗與挑戰，分享實用經驗與工具推薦。<br>
                    <a href='https://blog.twman.org/2024/08/LLM.html' target='_blank'>白話文手把手帶你科普 GenAI</a>：淺顯介紹生成式人工智慧核心概念，強調硬體資源和數據的重要性。<br>
                    <a href='https://blog.twman.org/2024/09/LLM.html' target='_blank'>大型語言模型直接就打完收工？</a>：回顧 LLM 領域探索歷程，討論硬體升級對 AI 開發的重要性。<br>
                    <a href='https://blog.twman.org/2024/07/RAG.html' target='_blank'>那些檢索增強生成要踩的坑</a>：探討 RAG 技術應用與挑戰，提供實用經驗分享和工具建議。<br>
                    <a href='https://blog.twman.org/2024/02/LLM.html' target='_blank'>那些大型語言模型要踩的坑</a>：探討多種 LLM 工具的應用與挑戰，強調硬體資源的重要性。<br>
                    <a href='https://blog.twman.org/2023/04/GPT.html' target='_blank'>Large Language Model，LLM</a>：探討 LLM 的發展與應用，強調硬體資源在開發中的關鍵作用。。<br>
                    <a href='https://blog.twman.org/2024/11/diffusion.html' target='_blank'>ComfyUI + Stable Diffuision</a>：深入探討影像生成與分割技術的應用，強調硬體資源的重要性。<br>
                    <a href='https://blog.twman.org/2024/02/asr-tts.html' target='_blank'>那些ASR和TTS可能會踩的坑</a>：探討 ASR 和 TTS 技術應用中的問題，強調數據質量的重要性。<br>
                    <a href='https://blog.twman.org/2021/04/NLP.html' target='_blank'>那些自然語言處理 (Natural Language Processing, NLP) 踩的坑</a>：分享 NLP 領域的實踐經驗，強調數據質量對模型效果的影響。<br>
                    <a href='https://blog.twman.org/2021/04/ASR.html' target='_blank'>那些語音處理 (Speech Processing) 踩的坑</a>：分享語音處理領域的實務經驗，強調資料品質對模型效果的影響。<br>
                    <a href='https://blog.twman.org/2023/07/wsl.html' target='_blank'>用PPOCRLabel來幫PaddleOCR做OCR的微調和標註</a><br>
                    <a href='https://blog.twman.org/2023/07/HugIE.html' target='_blank'>基於機器閱讀理解和指令微調的統一信息抽取框架之診斷書醫囑資訊擷取分析</a><br>                  
                <a href='https://github.com/facebookresearch/denoiser' target='_blank'> Real Time Speech Enhancement in the Waveform Domain (Interspeech 2020)</a>""",
    allow_flagging="never",
    examples=[
        ["exampleAudio/15s_2020-03-27_sep1.wav"],
        ["exampleAudio/13s_2020-03-27_sep2.wav"],
        ["exampleAudio/30s_2020-04-23_sep1.wav"],
        ["exampleAudio/15s_2020-04-23_sep2.wav"],
    ],
)

demo.launch(debug=True)