Wake Word Detection

# Path Configuration
BASE_PATH = "/home/daniel-de-anda/Escritorio/proyectos/make word"
PATH_POSITIVE = os.path.join(BASE_PATH, "audios")
PATH_NEGATIVE = os.path.join(BASE_PATH, "audios bad")
MODEL_SAVE_PATH = os.path.join(BASE_PATH, "wake_word_model.h5")

# Audio Parameters
SAMPLE_RATE = 16000  # Hz
DURATION = 5         # seconds (unified duration)
CHANNELS = 1         # Mono audio
N_MFCC = 13          # Number of MFCC coefficients

def extract_features(audio_data):
    """Extrae coeficientes MFCC de un array de audio."""
    # Ensure constant length (pad or trim)
    target_len = int(SAMPLE_RATE * DURATION)
    if len(audio_data) < target_len:
        audio_data = np.pad(audio_data, (0, target_len - len(audio_data)))
    else:
        audio_data = audio_data[:target_len]
    
    # Extract MFCC features
    mfcc = librosa.feature.mfcc(y=audio_data, sr=SAMPLE_RATE, n_mfcc=N_MFCC)
    return mfcc.T  # Shape: (time_steps, n_mfcc)

def prepare_dataset():
    x, y = [], []
    
    for path, label in [(PATH_POSITIVE, 1), (PATH_NEGATIVE, 0)]:
        print(f"📦 Cargando audios {'positivos' if label==1 else 'negativos'}...")
        if not os.path.exists(path):
            print(f"⚠️ Alerta: La carpeta {path} no existe.")
            continue
            
        for f in os.listdir(path):
            if f.endswith('.wav'):
                file_path = os.path.join(path, f)
                audio, _ = librosa.load(file_path, sr=SAMPLE_RATE, duration=DURATION)
                feat = extract_features(audio)
                x.append(feat)
                y.append(label)
                
    return np.array(x), np.array(y)

def build_model(input_shape):
    model = models.Sequential([
        layers.Input(shape=input_shape),
        layers.LSTM(64, return_sequences=False),
        layers.Dropout(0.3),
        layers.Dense(32, activation='relu'),
        layers.Dense(1, activation='sigmoid')
    ])
    model.compile(
        optimizer='adam', 
        loss='binary_crossentropy', 
        metrics=['accuracy']
    )
    return model

model.fit(
    x_train, 
    y_train, 
    epochs=40, 
    batch_size=8, 
    validation_split=0.2
)

def main():
    respuesta = input("¿Quieres entrenar el modelo? (s/n): ").lower()
    
    if respuesta == "s":
        # Training mode
        x_train, y_train = prepare_dataset()
        if len(x_train) == 0:
            print("❌ No hay datos suficientes.")
            return

        print(f"\n🧠 Entrenando con {len(x_train)} muestras...")
        model = build_model(x_train[0].shape)
        model.fit(x_train, y_train, epochs=40, batch_size=8, validation_split=0.2)
        model.save(MODEL_SAVE_PATH)
        print(f"✅ Modelo guardado en: {MODEL_SAVE_PATH}")
    else:
        # Load existing model
        if os.path.exists(MODEL_SAVE_PATH):
            print(f"✅ Cargando modelo existente...")
            model = models.load_model(MODEL_SAVE_PATH)
        else:
            print("❌ No hay modelo para cargar. Debes entrenar primero.")
            return

    # Detection loop (continues below)

# Detection configuration
print("\n" + "="*40)
print("🎤 MODO DETECCIÓN (Escuchando en bloques de 1s)")
print("="*40)

threshold = 0.8  # Confidence threshold

try:
    while True:
        # Record 5 seconds to match training data
        rec = sd.rec(
            int(DURATION * SAMPLE_RATE), 
            samplerate=SAMPLE_RATE, 
            channels=CHANNELS
        )
        sd.wait()
        audio_chunk = rec.flatten()
        
        # Extract features and predict
        feat = extract_features(audio_chunk)
        feat = np.expand_dims(feat, axis=0)  # Batch dimension
        
        prediction = model.predict(feat, verbose=0)[0][0]
        
        if prediction > threshold:
            print(f"✅ ¡WAKE WORD DETECTADO! ({prediction:.2f})")
        else:
            print(f". (Prob: {prediction:.2f})", end="\r")
            
except KeyboardInterrupt:
    print("\nDeteniendo detector...")

pip install numpy sounddevice tensorflow librosa

$ python entrenar_wake.py
¿Quieres entrenar el modelo? (s/n): s

📦 Cargando audios positivos...
📦 Cargando audios negativos...

🧠 Entrenando con 120 muestras...
Epoch 1/40
12/12 [==============================] - 2s 156ms/step - loss: 0.6821 - accuracy: 0.5625 - val_loss: 0.6745 - val_accuracy: 0.6250
Epoch 2/40
12/12 [==============================] - 1s 95ms/step - loss: 0.6523 - accuracy: 0.6458 - val_loss: 0.6234 - val_accuracy: 0.7083
...
Epoch 40/40
12/12 [==============================] - 1s 92ms/step - loss: 0.0823 - accuracy: 0.9792 - val_loss: 0.1234 - val_accuracy: 0.9583

✅ Modelo guardado en: /home/daniel-de-anda/Escritorio/proyectos/make word/wake_word_model.h5

$ python entrenar_wake.py
¿Quieres entrenar el modelo? (s/n): n
✅ Cargando modelo existente...

========================================
🎤 MODO DETECCIÓN (Escuchando en bloques de 1s)
========================================

. (Prob: 0.12)
. (Prob: 0.08)
. (Prob: 0.15)
✅ ¡WAKE WORD DETECTADO! (0.92)
. (Prob: 0.23)
. (Prob: 0.11)

target_len = 16000 * 5 = 80000 samples
MFCC shape after extraction: (time_steps, 13)
# time_steps depends on librosa's hop_length (default 512)
time_steps ≈ 80000 / 512 ≈ 156 frames

Final input shape: (156, 13)

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
lstm (LSTM)                  (None, 64)                19968     
_________________________________________________________________
dropout (Dropout)            (None, 64)                0         
_________________________________________________________________
dense (Dense)                (None, 32)                2080      
_________________________________________________________________
dense_1 (Dense)              (None, 1)                 33        
=================================================================
Total params: 22,081
Trainable params: 22,081
Non-trainable params: 0
_________________________________________________________________