import gradio as gr import base64 import json import numpy as np import torch import librosa import spaces # <-- Required for Hugging Face ZeroGPU infrastructure from fastapi import FastAPI, WebSocket, Request from fastapi.responses import HTMLResponse from transformers import AutoFeatureExtractor, Wav2Vec2ForSequenceClassification # ========================================== # 1. LOAD MODEL & FEATURE EXTRACTOR # ========================================== MODEL_ID = "Bgeorge/deepfake-audio-detector-wav2vec2" device = "cuda" if torch.cuda.is_available() else "cpu" print(f"📥 Loading AI Model ({MODEL_ID}) onto {device}...") feature_extractor = AutoFeatureExtractor.from_pretrained(MODEL_ID) model = Wav2Vec2ForSequenceClassification.from_pretrained(MODEL_ID).to(device) model.eval() # Global list to track active Flutter dashboard client connections connected_flutter_clients = [] # ========================================== # 2. ZERO-GPU WRAPPED INFERENCE FUNCTION # ========================================== @spaces.GPU(duration=30) def run_model_inference(audio_buffer): """Executes transformer inference securely under Hugging Face's ZeroGPU dynamic allocator.""" inputs = feature_extractor(audio_buffer, sampling_rate=16000, return_tensors="pt", padding=True) inputs = {k: v.to(device) for k, v in inputs.items()} with torch.no_grad(): outputs = model(**inputs) probs = torch.softmax(outputs.logits, dim=-1).squeeze().tolist() return probs # ========================================== # 3. GRADIO WEB INTERFACE (For Browser Testing) # ========================================== def predict_audio_file(audio_file): if audio_file is None: return "No audio uploaded.", 0.0 sr, y = audio_file if y.ndim > 1: y = np.mean(y, axis=1) # Convert stereo to mono y = y.astype(np.float32) # Resample if browser audio isn't 16kHz if sr != 16000: y = librosa.resample(y, orig_sr=sr, target_sr=16000) probs = run_model_inference(y) # Label 0 = REAL, Label 1 = FAKE fake_prob = probs[1] if isinstance(probs, list) and len(probs) > 1 else probs fake_confidence = round(fake_prob * 100, 2) real_confidence = round((1.0 - fake_prob) * 100, 2) status = "🚨 FAKE / SYNTHETIC VOICE" if fake_prob < 0.60 else "✅ REAL / HUMAN VOICE" score = fake_confidence if fake_prob > 0.60 else real_confidence return f"{status} | Confidence Score: {score}%", score with gr.Blocks(theme=gr.themes.Soft()) as demo: gr.Markdown("# 🛡️ VoiceGuard AI: Live Telephony Deepfake Shield") gr.Markdown("Test audio files via browser below. Live phone calls connect automatically through background Twilio webhooks.") with gr.Row(): audio_input = gr.Audio(sources=["upload", "microphone"], type="numpy", label="Upload Voice Sample") with gr.Row(): output_text = gr.Textbox(label="Classification Result") output_score = gr.Number(label="Confidence Score (%)") analyze_btn = gr.Button("Analyze Audio", variant="primary") analyze_btn.click(fn=predict_audio_file, inputs=audio_input, outputs=[output_text, output_score]) # ========================================== # 4. TWILIO & FLUTTER BACKEND (FastAPI) # ========================================== app = demo.app @app.post("/incoming-call") async def incoming_call(request: Request): """Twilio hits this webhook when an incoming call hits your phone number.""" host = request.url.hostname ws_url = f"wss://{host}/media-stream" twiml_response = f""" Security system active. Analyzing voice biometrics in real time. """ return HTMLResponse(content=twiml_response, media_type="application/xml") @app.websocket("/media-stream") async def handle_media_stream(websocket: WebSocket): """Twilio streams live raw call audio chunks here over WebSockets.""" await websocket.accept() print("📞 LIVE CALL CONNECTED: Streaming carrier audio...") audio_buffer = np.array([], dtype=np.float32) EVAL_INTERVAL = 16000 * 2 # Evaluate every 2 seconds of accumulated audio try: while True: message = await websocket.receive_text() data = json.loads(message) if data.get("event") == "media": chunk_bytes = base64.b64decode(data["media"]["payload"]) pcm_chunk = np.frombuffer(chunk_bytes, dtype=np.int16).astype(np.float32) / 32768.0 audio_16k = librosa.resample(pcm_chunk, orig_sr=8000, target_sr=16000) audio_buffer = np.concatenate([audio_buffer, audio_16k]) # Run sliding window evaluation when buffer hits interval threshold if len(audio_buffer) >= EVAL_INTERVAL: probs = run_model_inference(audio_buffer) fake_prob = probs[1] if isinstance(probs, list) and len(probs) > 1 else probs fake_confidence = round(fake_prob * 100, 2) real_confidence = round((1.0 - fake_prob) * 100, 2) is_fake = fake_prob > 0.60 score = fake_confidence if is_fake else real_confidence if is_fake: print(f"🚨 [LIVE FRAUD DETECTED] Synthetic Voice! Confidence: {score}%") else: print(f"✅ [CLEAN CALL] Real Human Voice. Confidence: {score}%") # Construct telemetry payload for Flutter dashboard telemetry_payload = json.dumps({ "is_fake": is_fake, "confidence_score": score }) # Broadcast telemetry to all connected Flutter simulator clients for client in list(connected_flutter_clients): try: await client.send_text(telemetry_payload) except Exception: connected_flutter_clients.remove(client) # Slide window forward: keep last 0.5 seconds for context overlap OVERLAP = int(16000 * 0.5) audio_buffer = audio_buffer[-OVERLAP:] elif data.get("event") == "stop": print("📞 Call hung up.") break except Exception as e: print(f"WebSocket error: {e}") @app.websocket("/ui-dashboard-feed") async def flutter_dashboard_feed(websocket: WebSocket): """WebSocket endpoint that connects your online Flutter simulator to receive live telemetry.""" await websocket.accept() connected_flutter_clients.append(websocket) print("📱 Flutter Dashboard Client Connected!") try: while True: await websocket.receive_text() except Exception: if websocket in connected_flutter_clients: connected_flutter_clients.remove(websocket) print("📱 Flutter Dashboard Client Disconnected.") # ========================================== # 5. LAUNCH APPLICATION # ========================================== if __name__ == "__main__": demo.launch(server_name="0.0.0.0", server_port=7860)