app.py

import gradio as gr
import torch
import torchaudio
import torchaudio.transforms as T
import traceback
import matplotlib.pyplot as plt
import io
from PIL import Image

# Ensure AudioSeal is imported correctly
try:
    from audioseal import AudioSeal
    print("AudioSeal is available for watermark detection.")
except ImportError as e:
    print(f"AudioSeal could not be imported: {e}")

def load_and_resample_audio(audio_file_path, target_sample_rate=16000):
    waveform, sample_rate = torchaudio.load(audio_file_path)
    if sample_rate != target_sample_rate:
        resampler = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=target_sample_rate)
        waveform = resampler(waveform)
    return waveform, target_sample_rate

def extract_mfcc_features(waveform, sample_rate, n_mfcc=40, n_mels=128, win_length=400, hop_length=160):
    mfcc_transform = T.MFCC(
        sample_rate=sample_rate,
        n_mfcc=n_mfcc,
        melkwargs={
            'n_fft': 400,
            'n_mels': n_mels,
            'hop_length': hop_length,
            'win_length': win_length
        }
    )
    mfcc = mfcc_transform(waveform)
    return mfcc.mean(dim=2)
    
def plot_spectrogram(waveform, sample_rate):
    if waveform.ndim == 1:
        waveform = waveform.unsqueeze(0)  # Ensure waveform is 2D
    spectrogram_transform = T.Spectrogram()
    spectrogram = spectrogram_transform(waveform)
    spectrogram_db = torchaudio.transforms.AmplitudeToDB()(spectrogram)
    plt.figure(figsize=(10, 4))
    plt.imshow(spectrogram_db[0].numpy(), cmap='hot', aspect='auto', origin='lower')
    plt.axis('off')  # Hide axes for a clean image
    buf = io.BytesIO()
    plt.savefig(buf, format='png', bbox_inches='tight', pad_inches=0)
    plt.close()
    buf.seek(0)
    return Image.open(buf)

def detect_watermark(audio_data, sample_rate):
    # Ensure AudioSeal is available
    if 'AudioSeal' not in globals():
        spectrogram_image = plot_spectrogram(audio_data, sample_rate)
        return "AudioSeal not available", spectrogram_image
    
    # Load audio data correctly
    waveform, sr = load_and_resample_audio(audio_data, target_sample_rate=16000)
    
    # Ensure waveform is a tensor before passing to the detector
    if not isinstance(waveform, torch.Tensor):
        return "Error: waveform is not a tensor.", plot_spectrogram(waveform, sr)
    
    # Load the detector
    detector = AudioSeal.load_detector("audioseal_detector_16bits")
    
    # Process waveform with the detector
    results, messages = detector.forward(waveform.unsqueeze(0), sample_rate=sample_rate)  # Ensure waveform is in batch form
    detect_probs = results[:, 1, :]
    result = detect_probs.mean().cpu().item()
    message = f"Detection result: {'Watermarked Audio' if result > 0.5 else 'Not watermarked'}"
    spectrogram_image = plot_spectrogram(waveform, sr)
    
    return message, spectrogram_image

def main(audio_file_path):
    waveform, resampled_sr = load_and_resample_audio(audio_file_path)
    plot_spectrogram(waveform, resampled_sr)
    samples_per_batch = 5 * resampled_sr  # 5s audios
    audio_batches = torch.split(waveform, samples_per_batch, dim=1)[:-1]  # Exclude the last batch if it's not 5 seconds long
    audio_batched = torch.concat(audio_batches, dim=0)
    audio_batched = audio_batched.unsqueeze(1) # add channel dimension
    result = detect_watermark(audio_batched, resampled_sr)
    print(f"Probability of watermark: {result}")

# Gradio interface
interface = gr.Interface(
    fn=detect_watermark,
    inputs=gr.Audio(label="Upload your audio", type="filepath"),
    outputs=["text", "image"],
    title="Deep Fake Defender: AudioSeal Watermark Detection",
    description="Analyzes audio to detect AI-generated content."
)

if __name__ == "__main__":
    interface.launch()