import gradio as gr
import librosa
import numpy as np
import tensorflow as tf
from tensorflow.keras.models import load_model

# Load your pre-trained model (make sure it's in the same directory or provide a path)
model = load_model('voice_authentication_model.keras')  # Replace with your model's filename

# Function to extract MFCC features and make a prediction
def predict_user_or_non_user(audio):
    try:
        # Load the audio file using librosa
        y, sr = librosa.load(audio, sr=16000)  # sr=None to keep the original sampling rate
        
        # Optional: Normalize the audio volume (if necessary)
        y = librosa.util.normalize(y)
        
        # Extract MFCC features from the audio
        mfccs = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=13)  # You can adjust n_mfcc as needed
        mfccs = np.mean(mfccs.T, axis=0)  # Take the mean of MFCCs over time to reduce dimension
        
        # Reshape the MFCCs to match the input shape expected by the model
        mfccs = mfccs.reshape(1, -1)  # Reshape to 1 sample, with the number of features
        
        # Predict the class (user or non-user)
        prediction = model.predict(mfccs)
        
        # Debugging: print raw model output
        print(f"Raw Prediction Output: {prediction}")
        
        # Apply thresholding based on the raw prediction value
        # If the model outputs a probability, try adjusting the threshold (e.g., 0.6 instead of 0.5)
        if prediction[0] > 0.5:
            return "User"
        else:
            return "Non-User"
    except Exception as e:
        print(f"Error in prediction: {e}")
        return "Error during prediction"

# Define the Gradio interface
iface = gr.Interface(
    fn=predict_user_or_non_user,  # The function to call when an audio input is given
    inputs=gr.Audio(type="filepath"),  # Corrected audio input setup
    outputs="text",  # Output will be text (User or Non-user)
    live=True  # Live mode so the interface updates in real-time
)

# Launch the interface
iface.launch()