app.py

import gradio as gr
import torch
import torchaudio
from torchaudio.transforms import Resample
from transformers import AutoFeatureExtractor, AutoModelForAudioClassification

# Load the HF feature extractor and model
feature_extractor = AutoFeatureExtractor.from_pretrained(
    "MelodyMachine/Deepfake-audio-detection-V2"
)
model = AutoModelForAudioClassification.from_pretrained(
    "MelodyMachine/Deepfake-audio-detection-V2"
)

TARGET_SR = feature_extractor.sampling_rate  # should be 16000

def detect_deepfake_audio(audio_path: str) -> str:
    # Load audio file
    waveform, orig_sr = torchaudio.load(audio_path)

    # Mix to mono if necessary
    if waveform.shape[0] > 1:
        waveform = torch.mean(waveform, dim=0, keepdim=True)

    # Resample if not already 16 kHz
    if orig_sr != TARGET_SR:
        resampler = Resample(orig_sr, TARGET_SR)
        waveform = resampler(waveform)

    # Prepare inputs
    inputs = feature_extractor(
        waveform, sampling_rate=TARGET_SR, return_tensors="pt"
    )
    with torch.no_grad():
        outputs = model(**inputs)

    # Compute probabilities
    probs = torch.softmax(outputs.logits, dim=-1)[0]
    idx = torch.argmax(probs).item()
    label = model.config.id2label[idx]
    confidence = probs[idx].item()

    return f"The audio is classified as **{label}** with confidence **{confidence:.2f}**"

# Build the Gradio Blocks interface
with gr.Blocks() as demo:
    gr.Markdown("# Audio Deepfake Detection")
    gr.Markdown("Upload an audio clip to check for deepfake content.")
    audio_in = gr.Audio(type="filepath", label="Select Audio File")
    txt_out = gr.Textbox(label="Result")
    gr.Button("Detect").click(
        fn=detect_deepfake_audio, inputs=audio_in, outputs=txt_out
    )

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