app.py

import gradio as gr
import librosa
import numpy as np
import torch
import matplotlib.pyplot as plt
from transformers import AutoModelForAudioClassification, ASTFeatureExtractor
import random
import tempfile

# Model and feature extractor loading
model = AutoModelForAudioClassification.from_pretrained("./")
feature_extractor = ASTFeatureExtractor.from_pretrained("./")

def plot_waveform(waveform, sr):
    plt.figure(figsize=(12, 4))
    plt.title('Waveform')
    plt.ylabel('Amplitude')
    plt.plot(np.linspace(0, len(waveform) / sr, len(waveform)), waveform)
    plt.xlabel('Time (s)')
    # Save plot to a temporary file
    temp_file = tempfile.NamedTemporaryFile(delete=False, suffix='.png', dir='./')
    plt.savefig(temp_file.name)
    plt.close()  # Close the figure to free memory
    return temp_file.name

def plot_spectrogram(waveform, sr):
    S = librosa.feature.melspectrogram(y=waveform, sr=sr, n_mels=128)
    S_DB = librosa.power_to_db(S, ref=np.max)
    plt.figure(figsize=(12, 6))
    librosa.display.specshow(S_DB, sr=sr, x_axis='time', y_axis='mel')
    plt.title('Mel Spectrogram')
    plt.colorbar(format='%+2.0f dB')
    plt.tight_layout()
    # Save plot to a temporary file
    temp_file = tempfile.NamedTemporaryFile(delete=False, suffix='.png', dir='./')
    plt.savefig(temp_file.name)
    plt.close()  # Close the figure to free memory
    return temp_file.name

def custom_feature_extraction(audio, sr=16000, target_length=1024):
    features = feature_extractor(audio, sampling_rate=sr, return_tensors="pt", padding="max_length", max_length=target_length)
    return features.input_values

def apply_time_shift(waveform, max_shift_fraction=0.1):
    shift = random.randint(-int(max_shift_fraction * len(waveform)), int(max_shift_fraction * len(waveform)))
    return np.roll(waveform, shift)

def predict_voice(audio_file_path):
    try:
        waveform, sample_rate = librosa.load(audio_file_path, sr=feature_extractor.sampling_rate, mono=True)
        augmented_waveform = apply_time_shift(waveform)
        
        original_features = custom_feature_extraction(waveform, sr=sample_rate)
        augmented_features = custom_feature_extraction(augmented_waveform, sr=sample_rate)
        
        with torch.no_grad():
            outputs_original = model(original_features)
            outputs_augmented = model(augmented_features)
        
        logits = (outputs_original.logits + outputs_augmented.logits) / 2
        predicted_index = logits.argmax()
        original_label = model.config.id2label[predicted_index.item()]
        confidence = torch.softmax(logits, dim=1).max().item() * 100
        
        # Map original labels to new labels
        label_mapping = {
            "Spoof": "AI-generated Clone",
            "Bonafide": "Real Human Voice"
        }
        # Use the original label to get the new label
        new_label = label_mapping.get(original_label, "Unknown")  # Default to "Unknown" if label not found
        
        waveform_plot = plot_waveform(waveform, sample_rate)
        spectrogram_plot = plot_spectrogram(waveform, sample_rate)
        
        return (
            f"The voice is classified as '{new_label}' with a confidence of {confidence:.2f}%.",
            waveform_plot,
            spectrogram_plot
        )
    except Exception as e:
        return f"Error during processing: {e}", None, None

iface = gr.Interface(
    fn=predict_voice,
    inputs=gr.Audio(label="Upload Audio File", type="filepath"),
    outputs=[
        gr.Textbox(label="Prediction"),
        gr.Image(label="Waveform"),  # Adjusted to remove unsupported 'tool' argument
        gr.Image(label="Spectrogram")  # Adjusted to remove unsupported 'tool' argument
    ],
    title="Voice Clone Detection",
    description="Detects whether a voice is real or an AI-generated clone. Upload an audio file to see the results."
)

iface.launch()