Update app.py

app.py

@@ -12,32 +12,35 @@ CORS(app)
 MODEL_PATH = "CRNN_model_final.h5"
 
 print("Chargement du modele CRNN...")
-model = tf.keras.models.load_model(MODEL_PATH)
+model = tf.keras.models.load_model(MODEL_PATH, compile=False)
 print("Modele charge!")
 print(f"Input shape: {model.input_shape}")
 print(f"Output shape: {model.output_shape}")
 
-def extract_mfcc(audio, sr, n_mfcc=13):
-    mfccs = librosa.feature.mfcc(y=audio, sr=sr, n_mfcc=n_mfcc)
-    mfccs = (mfccs - np.mean(mfccs)) / (np.std(mfccs) + 1e-8)
-    return mfccs
+SILENCE_CLASSES = [13, 14]
 
 def predict_vad(audio, sr):
-    mfccs = extract_mfcc(audio, sr, n_mfcc=13)
-    features = mfccs.T
-    features = np.expand_dims(features, axis=0)
+    if sr != 16000:
+        audio = librosa.resample(audio, orig_sr=sr, target_sr=16000)
+    
+    target_length = 1024
+    
+    if len(audio) < target_length:
+        audio = np.pad(audio, (0, target_length - len(audio)), mode='constant')
+    elif len(audio) > target_length:
+        start = (len(audio) - target_length) // 2
+        audio = audio[start:start + target_length]
+    
+    audio = audio / (np.max(np.abs(audio)) + 1e-8)
+    features = audio.reshape(1, target_length, 1)
     
     prediction = model.predict(features, verbose=0)
+    predicted_class = int(np.argmax(prediction[0]))
+    confidence = float(np.max(prediction[0]))
     
-    if prediction.shape[-1] == 2:
-        confidence = float(np.max(prediction))
-        label = "voice" if np.argmax(prediction) == 1 else "silence"
-    else:
-        confidence = float(prediction[0][0])
-        label = "voice" if confidence > 0.5 else "silence"
-        confidence = confidence if label == "voice" else 1 - confidence
+    label = "silence" if predicted_class in SILENCE_CLASSES else "voice"
     
-    return label, confidence
+    return label, confidence, predicted_class
 
 @app.after_request
 def after_request(response):
@@ -54,7 +57,7 @@ def home():
 def health():
     if request.method == 'OPTIONS':
         return make_response('', 204)
-    return jsonify({'status': 'ok', 'model': 'CRNN VAD + GCC-PHAT DOA'})
+    return jsonify({'status': 'ok', 'model': 'CRNN VAD'})
 
 @app.route('/api/analyze', methods=['POST', 'OPTIONS'])
 def analyze():
@@ -75,38 +78,39 @@ def analyze():
         
         n_channels = y.shape[0] if y.ndim > 1 else 1
         duration = y.shape[-1] / sr
-        n_frames = max(1, int(duration * 10))
         
-        vad_label, vad_confidence = predict_vad(y_mono, sr)
+        vad_label, vad_confidence, pred_class = predict_vad(y_mono, sr)
+        
+        frame_size = 1024
+        n_frames = max(1, len(y_mono) // frame_size)
         
-        frame_length = len(y_mono) // n_frames
         vad_timeline = []
         for i in range(n_frames):
-            start_sample = i * frame_length
-            end_sample = min((i + 1) * frame_length, len(y_mono))
+            start_sample = i * frame_size
+            end_sample = min(start_sample + frame_size, len(y_mono))
             frame_audio = y_mono[start_sample:end_sample]
             
-            if len(frame_audio) > sr * 0.05:
-                frame_label, frame_conf = predict_vad(frame_audio, sr)
+            if len(frame_audio) >= 256:
+                frame_label, frame_conf, _ = predict_vad(frame_audio, sr)
             else:
                 frame_label, frame_conf = vad_label, vad_confidence
             
             vad_timeline.append({
-                'start': round(i * 0.1, 2),
-                'end': round((i + 1) * 0.1, 2),
+                'start': round(start_sample / sr, 3),
+                'end': round(end_sample / sr, 3),
                 'label': frame_label,
                 'confidence': round(frame_conf, 3)
             })
         
         doa_angle = -25.5 + np.random.randn() * 10
-        doa_trajectory = [{'time': round(i*0.1,2), 'angle': round(doa_angle + np.random.randn()*3, 1)} for i in range(n_frames)]
+        doa_trajectory = [{'time': round(i * (duration/n_frames), 2), 'angle': round(doa_angle + np.random.randn()*3, 1)} for i in range(n_frames)]
         
         voice_ratio = sum(1 for f in vad_timeline if f['label'] == 'voice') / len(vad_timeline)
         
         return jsonify({
             'success': True,
             'metadata': {'channels': int(n_channels), 'sampleRate': int(sr), 'duration': round(duration, 2), 'samples': int(y.shape[-1])},
-            'vad': {'prediction': vad_label, 'confidence': round(vad_confidence, 3), 'timeline': vad_timeline},
+            'vad': {'prediction': vad_label, 'confidence': round(vad_confidence, 3), 'predicted_class': pred_class, 'timeline': vad_timeline},
             'doa': {'angle': round(doa_angle, 1), 'confidence': 0.89, 'trajectory': doa_trajectory},
             'metrics': {'snr': round(10 * np.log10(np.mean(y_mono**2) / 1e-10), 1), 'voiceRatio': round(voice_ratio, 2), 'meanDoa': round(np.mean([d['angle'] for d in doa_trajectory]), 1)}
         })