app.py

import gradio as gr
import librosa
import numpy as np
import torch
import logging
from transformers import AutoModelForAudioClassification

logging.basicConfig(level=logging.INFO)

model_path = "./"
model = AutoModelForAudioClassification.from_pretrained(model_path)

def preprocess_audio(audio_file_path, sr=16000):
    waveform, _ = librosa.load(audio_file_path, sr=sr)
    waveform = librosa.effects.trim(waveform)[0]  # Trim silence
    return waveform

def extract_features(waveform, sr=16000, n_mels=128, n_fft=2048, hop_length=512):
    S = librosa.feature.melspectrogram(y=waveform, sr=sr, n_mels=n_mels, n_fft=n_fft, hop_length=hop_length)
    S_DB = librosa.power_to_db(S, ref=np.max)
    return torch.tensor(S_DB).float().unsqueeze(0)  # Add batch dimension

def predict_voice(audio_file_path):
    try:
        waveform = preprocess_audio(audio_file_path)
        features = extract_features(waveform)
        
        with torch.no_grad():
            outputs = model(features)
            logits = outputs.logits
            predicted_index = logits.argmax()
            label = model.config.id2label[predicted_index.item()]
            confidence = torch.softmax(logits, dim=1).max().item() * 100
        
        result = f"The voice is classified as '{label}' with a confidence of {confidence:.2f}%."
        logging.info("Prediction successful.")
    except Exception as e:
        result = f"Error during processing: {e}"
        logging.error(result)

    return result

iface = gr.Interface(
    fn=predict_voice,
    inputs=gr.Audio(label="Upload Audio File", type="file"),
    outputs=gr.Text(label="Prediction"),
    title="Voice Authenticity Detection",
    description="Detects whether a voice is real or AI-generated. Upload an audio file to see the results."
)

iface.launch()