app.py

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

# Load the model directly from the .h5 file
model_path = os.path.join(os.path.dirname(__file__), 'wav2vec_model.h5')
model = tf.keras.models.load_model(model_path)

# Define emotions list
emotions = ["neutral", "happy", "sad", "angry", "fearful", "disgust", "surprised"]

def extract_features(audio_path, sample_rate=16000, n_mfcc=13, max_length=128):
    """Extract MFCC features from an audio file"""
    try:
        audio, sr = librosa.load(audio_path, sr=sample_rate)
        
        # Extract MFCCs
        mfccs = librosa.feature.mfcc(y=audio, sr=sr, n_mfcc=n_mfcc)
        
        # Pad or truncate to fixed length
        if mfccs.shape[1] < max_length:
            pad_width = max_length - mfccs.shape[1]
            mfccs = np.pad(mfccs, pad_width=((0, 0), (0, pad_width)), mode='constant')
        else:
            mfccs = mfccs[:, :max_length]
        
        return mfccs
    except Exception as e:
        print(f"Error in feature extraction: {e}")
        return None

def predict_emotion(audio):
    """Predict emotion from audio input
    
    This function accepts both file path (when uploading) and audio array 
    (when recording via microphone) as input
    """
    try:
        # Check if audio is a file path or audio array
        if isinstance(audio, str):  # File path
            features = extract_features(audio)
        else:  # Audio array from microphone
            # If audio is a tuple (audio array, sample rate)
            if isinstance(audio, tuple):
                audio_array, sample_rate = audio
            else:
                # If only audio array is provided, assume sample rate
                audio_array = audio
                sample_rate = 16000
                
            # Convert to mono if stereo
            if len(audio_array.shape) > 1:
                audio_array = np.mean(audio_array, axis=1)
                
            # Extract features
            mfccs = librosa.feature.mfcc(y=audio_array, sr=sample_rate, n_mfcc=13)
            
            # Pad or truncate to fixed length
            max_length = 128
            if mfccs.shape[1] < max_length:
                pad_width = max_length - mfccs.shape[1]
                mfccs = np.pad(mfccs, pad_width=((0, 0), (0, pad_width)), mode='constant')
            else:
                mfccs = mfccs[:, :max_length]
                
            features = mfccs
        
        if features is None:
            return {emotion: 0.0 for emotion in emotions}
        
        # Reshape for model input
        features = np.expand_dims(features, axis=0)
        
        # Make prediction
        predictions = model.predict(features)
        
        # Format results
        result = {emotion: float(predictions[0][i]) for i, emotion in enumerate(emotions)}
        return result
        
    except Exception as e:
        print(f"Error in prediction: {e}")
        return {emotion: 0.0 for emotion in emotions}

# Create Gradio interface with both file upload and microphone
demo = gr.Interface(
    fn=predict_emotion,
    inputs=gr.Audio(sources=["microphone", "upload"], type="filepath"),
    outputs=gr.Label(num_top_classes=7),
    title="Speech Emotion Recognition",
    description="Upload an audio file or record your voice to identify the emotion. This model can detect neutral, happy, sad, angry, fearful, disgust, and surprised emotions.",
    examples=[
        ["example1.wav"],  # Add example files here if you have them
    ]
)

demo.launch()