app.py

import gradio as gr
import tensorflow as tf
import librosa
import numpy as np
import tempfile

# Diccionario de etiquetas
labels = [
    'cat', 'house', 'marvin', 'nine', 'one', 'eight', 'three', 'five', 'zero', 'four',
    'tree', 'wow', 'six', 'off', 'bed', 'seven', 'down', 'happy', 'on', 'yes',
    'two', 'right', 'left', 'stop', 'go', 'no', 'sheila', 'up', 'bird', 'dog'
]


def extract_features(file_name):
    try:
        # Resamplea a 16kHz
        audio, sample_rate = librosa.load(file_name, sr=16000)
        
        # Saca Mel-spectrograma
        mel_spectrogram = librosa.feature.melspectrogram(
            y=audio, 
            sr=sample_rate, 
            n_mels=257, 
            n_fft=512, 
            hop_length=256
        )
        
        # Convierte a escala logarítmica
        log_mel_spectrogram = librosa.power_to_db(mel_spectrogram, ref=np.max)
        
        # Ajusta tamaño exacto
        log_mel_spectrogram = librosa.util.fix_length(log_mel_spectrogram, size=257, axis=0)
        log_mel_spectrogram = librosa.util.fix_length(log_mel_spectrogram, size=97, axis=1)
        
        # Normaliza
        log_mel_spectrogram = (log_mel_spectrogram - np.mean(log_mel_spectrogram)) / np.std(log_mel_spectrogram)

        # Añade canal
        log_mel_spectrogram = log_mel_spectrogram[..., np.newaxis]
        
    except Exception as e:
        print(f"Error encountered while parsing file: {file_name}")
        print(e)
        return None 
    
    return log_mel_spectrogram


def classify_audio(audio_file):
    print(f"Tipo de audio_file: {type(audio_file)}") 

    # El tipo es string (ruta), no hace falta leer ni escribir en temp files
    file_path = audio_file

    # Extrae características
    features = extract_features(file_path)

    if features is None:
        return "Error al procesar el audio"

    # Añade batch dimension
    features = features[np.newaxis, ...]  # (1, 97, 257, 1)

    # Carga el modelo en CPU
    model = tf.keras.models.load_model('my_model.h5', compile=False)

    with tf.device('/CPU:0'):
        prediction = model.predict(features)
        predicted_label_index = np.argmax(prediction)

    predicted_label = labels[predicted_label_index]
    return predicted_label

iface = gr.Interface(
    fn=classify_audio,
    inputs=gr.Audio(type="filepath"),
    outputs="text",
    title="Clasificación de audio simple",
    description="Sube un archivo de audio para clasificarlo."
)

iface.launch()