import os
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
# 統一音訊預處理流程
def preprocess_audio(input_path, output_path):
# 用 pydub 處理格式轉換
audio = AudioSegment.from_file(input_path)
# 強制轉單聲道 + 16kHz 採樣率
if audio.channels > 1:
audio = audio.set_channels(1)
if audio.frame_rate != 16000:
audio = audio.set_frame_rate(16000)
# 導出為 WAV 暫存檔
audio.export(output_path, format="wav")
return output_path
# 處理 MP3 特殊流程
if file.lower().endswith(".mp3"):
temp_wav = "temp_input.wav"
preprocess_audio(file, temp_wav)
file = temp_wav
# 載入模型
model = Demucs(hidden=64)
state_dict = torch.load(modelpath, map_location='cpu')
model.load_state_dict(state_dict)
model.eval()
# 載入音訊並強制轉單聲道
x, sr = torchaudio.load(file)
if x.shape[0] > 1:
x = torch.mean(x, dim=0, keepdim=True)
# 音訊長度檢查
MAX_AUDIO_SECONDS = 900
if x.shape[1] / sr > MAX_AUDIO_SECONDS:
raise ValueError(f"音訊過長！限制：{MAX_AUDIO_SECONDS} 秒，當前：{x.shape[1]/sr:.1f} 秒")
# 執行降噪
with torch.no_grad():
out = model(x[None])[0]
# 後處理
out = out / max(out.abs().max().item(), 1)
torchaudio.save('enhanced.wav', out, sr)
# 轉 MP3 輸出
enhanced_mp3 = 'enhanced.mp3'
AudioSegment.from_wav('enhanced.wav').export(
enhanced_mp3,
format="mp3",
bitrate="256k"
)
# 清理暫存檔
if os.path.exists("temp_input.wav"):
os.remove("temp_input.wav")
return enhanced_mp3
# 👇 重要：修正 Gradio 類型推導問題
transcribe.__annotations__ = {
"file_upload": str,
"microphone": str,
"return": str
}
# 🎯 你提供的 description 內容（已轉為 HTML）
description_html = """
上傳一段音檔 （支援 .mp3, .wav），為了提升語音識別的效果，可以在識別前先進行噪音去除
📘 相關技術文章：
"""
demo = gr.Interface(
fn=transcribe,
inputs=[
gr.Audio(type="filepath", label="上傳音訊檔案", sources=["upload", "microphone"])
],
outputs=[
gr.Audio(type="filepath", label="處理後音訊")
],
live=True,
allow_flagging="never",
description=description_html
)
demo.launch(debug=True, share=True)