Update app.py

app.py

@@ -4,51 +4,35 @@ import numpy as np
 import librosa
 import io
 import os
-import gdown
 import tensorflow as tf
 
 app = Flask(__name__)
 CORS(app)
 
-# Telecharger et charger le modele au demarrage
-MODEL_URL = "https://drive.google.com/uc?id=1_eUJwfSSab9bQFW5Ow4kL54OMr6cGaO0"
-MODEL_PATH = "/app/model.h5"
+MODEL_PATH = "best_crnn_model_accuracy.keras"
 
-print("Telechargement du modele CNN-1D...")
-gdown.download(MODEL_URL, MODEL_PATH, quiet=False)
-print("Chargement du modele...")
+print("Chargement du modele CRNN...")
 model = tf.keras.models.load_model(MODEL_PATH)
-print("Modele charge avec succes!")
+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):
-    """Extrait les MFCC du signal audio"""
     mfccs = librosa.feature.mfcc(y=audio, sr=sr, n_mfcc=n_mfcc)
-    # Normaliser
     mfccs = (mfccs - np.mean(mfccs)) / (np.std(mfccs) + 1e-8)
     return mfccs
 
 def predict_vad(audio, sr):
-    """Prediction VAD avec le modele CNN-1D"""
-    # Extraire MFCC
     mfccs = extract_mfcc(audio, sr, n_mfcc=13)
+    features = mfccs.T
+    features = np.expand_dims(features, axis=0)
     
-    # Adapter la forme pour le modele (ajuster selon votre modele)
-    # Shape typique: (batch, time_steps, n_mfcc) ou (batch, n_mfcc, time_steps)
-    features = mfccs.T  # (time_steps, n_mfcc)
-    features = np.expand_dims(features, axis=0)  # (1, time_steps, n_mfcc)
-    
-    # Prediction
     prediction = model.predict(features, verbose=0)
     
-    # Interpreter la sortie (ajuster selon votre modele)
     if prediction.shape[-1] == 2:
-        # Classification binaire avec softmax
         confidence = float(np.max(prediction))
         label = "voice" if np.argmax(prediction) == 1 else "silence"
     else:
-        # Classification binaire avec sigmoid
         confidence = float(prediction[0][0])
         label = "voice" if confidence > 0.5 else "silence"
         confidence = confidence if label == "voice" else 1 - confidence
@@ -64,13 +48,13 @@ def after_request(response):
 
 @app.route('/', methods=['GET'])
 def home():
-    return jsonify({'message': 'AcoustiTrack API', 'status': 'running', 'model': 'CNN-1D VAD'})
+    return jsonify({'message': 'AcoustiTrack API', 'status': 'running'})
 
 @app.route('/api/health', methods=['GET', 'OPTIONS'])
 def health():
     if request.method == 'OPTIONS':
         return make_response('', 204)
-    return jsonify({'status': 'ok', 'model': 'CNN-1D VAD + GCC-PHAT DOA'})
+    return jsonify({'status': 'ok', 'model': 'CRNN VAD + GCC-PHAT DOA'})
 
 @app.route('/api/analyze', methods=['POST', 'OPTIONS'])
 def analyze():
@@ -83,7 +67,6 @@ def analyze():
         audio_file = request.files['audio']
         y, sr = librosa.load(io.BytesIO(audio_file.read()), sr=16000, mono=False)
         
-        # Gerer mono/multi-canal
         if y.ndim == 1:
             y_mono = y
             y = np.tile(y, (4, 1))
@@ -94,10 +77,8 @@ def analyze():
         duration = y.shape[-1] / sr
         n_frames = max(1, int(duration * 10))
         
-        # Prediction VAD avec le modele CNN-1D
         vad_label, vad_confidence = predict_vad(y_mono, sr)
         
-        # Generer timeline VAD frame par frame
         frame_length = len(y_mono) // n_frames
         vad_timeline = []
         for i in range(n_frames):
@@ -105,11 +86,10 @@ def analyze():
             end_sample = min((i + 1) * frame_length, len(y_mono))
             frame_audio = y_mono[start_sample:end_sample]
             
-            if len(frame_audio) > sr * 0.05:  # Au moins 50ms
+            if len(frame_audio) > sr * 0.05:
                 frame_label, frame_conf = predict_vad(frame_audio, sr)
             else:
-                frame_label = vad_label
-                frame_conf = vad_confidence
+                frame_label, frame_conf = vad_label, vad_confidence
             
             vad_timeline.append({
                 'start': round(i * 0.1, 2),
@@ -118,7 +98,6 @@ def analyze():
                 'confidence': round(frame_conf, 3)
             })
         
-        # DOA estimation (simulation - remplacer par GCC-PHAT reel si disponible)
         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)]
         
@@ -126,27 +105,10 @@ def analyze():
         
         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
-            },
-            '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)
-            }
+            '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},
+            '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)}
         })
     except Exception as e:
         import traceback