Update app.py

app.py

@@ -9,43 +9,48 @@ from denoiser.demucs import Demucs
 from pydub import AudioSegment
 from transformers import AutoModelForSequenceClassification, AutoTokenizer
 
-# 設置 Hugging Face Hub 的 Access Token
-auth_token = os.getenv("HF_TOKEN")
+# # 設置 Hugging Face Hub 的 Access Token
+# auth_token = os.getenv("HF_TOKEN")
 
-# 加載私有模型
-model_id = "DeepLearning101/Speech-Quality-Inspection_Meta-Denoiser"
-model = AutoModelForSequenceClassification.from_pretrained(model_id, token=auth_token)
-tokenizer = AutoTokenizer.from_pretrained(model_id, token=auth_token)
+# # 加載私有模型
+# model_id = "DeepLearning101/Speech-Quality-Inspection_Meta-Denoiser"
+# model = AutoModelForSequenceClassification.from_pretrained(model_id, token=auth_token)
+# tokenizer = AutoTokenizer.from_pretrained(model_id, token=auth_token)
+
+modelpath = './denoiser/master64.th'
 
 def transcribe(file_upload, microphone):
     file = microphone if microphone is not None else file_upload
-    demucs_model = Demucs(hidden=64)
-    state_dict = torch.load("path_to_model_checkpoint", map_location='cpu')  # 請確保提供正確的模型文件路徑
-    demucs_model.load_state_dict(state_dict)
+    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_model(x[None])[0]
+    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 = 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"
 
-    # 假設模型是用於文本分類
-    inputs = tokenizer("enhanced.wav", return_tensors="pt")
-    outputs = model(**inputs)
-    predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
+    # # 假設模型是用於文本分類
+    # inputs = tokenizer("enhanced.wav", return_tensors="pt")
+    # outputs = model(**inputs)
+    # predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)
 
-    return "enhanced.wav", predictions
+    # return "enhanced.wav", predictions
 
 demo = gr.Interface(
     fn=transcribe,
     inputs=[
-        gr.Audio(type="filepath", label="語音質檢麥克風實時錄音"),
-        gr.Audio(type="filepath", label="語音質檢原始音檔"),
-    ],
-    outputs=[
-        gr.Audio(type="filepath", label="Output"),
-        gr.Textbox(label="Model Predictions")
+        gr.Audio(source="microphone", type="filepath", optional=True, label="語音質檢麥克風實時錄音"),        
+        gr.Audio(source="upload", type="filepath", optional=True, label="語音質檢原始音檔"),
     ],
+    outputs=gr.Audio(type="filepath", label="Output"),
+    # outputs=[
+    #     gr.Audio(type="filepath", label="Output"),
+    #     gr.Textbox(label="Model Predictions")
+    # ],
     title="<p style='text-align: center'><a href='https://www.twman.org/AI' target='_blank'>語音質檢噪音去除 (語音增強)：Meta Denoiser</a>",
     description="為了提升語音識別的效果，可以在識別前先進行噪音去除",
     allow_flagging="never",
@@ -57,4 +62,4 @@ demo = gr.Interface(
     ],
 )
 
-demo.launch(debug=True)
+demo.launch(enable_queue=True, debug=True)