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"

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="語音質檢/噪音去除 (語音增強)",
    description="""<a href='https://www.twman.org' target='_blank'>TonTon Huang Ph.D. @ 2024/04 </a><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://blog.twman.org/2021/04/ASR.html' target='_blank'>那些語音處理 (Speech Processing) 踩的坑</a> | <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><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> | <a href='https://blog.twman.org/2024/02/LLM.html' target='_blank'>那些大模型開發會踩的坑</a><br>
                <a href='https://blog.twman.org/2023/04/GPT.html' target='_blank'>什麼是大語言模型，它是什麼？想要嗎？</a><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> | <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>
                基於<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)