TrueVoice — Detecting AI Voice-Cloning Scams with Gemma 4
Winner — Safety & Trust Track
Built for the Gemma 4 Good Hackathon (Google DeepMind × Kaggle; $200K prize pool, 5 impact tracks) as a 2-person team, TrueVoice won the Safety & Trust track.
Why This Matters
Voice cloning has legitimate uses, but the barrier to misuse is collapsing: a convincing clone now needs only a few seconds of public audio and a free tool. AI-powered voice-phishing caused over $5M in documented US losses in 2025, and the FBI attributes hundreds of millions more to AI-enabled scams. Enterprise anti-spoofing systems are expensive and API-locked; research systems aren’t deployable. TrueVoice fills that gap — a browser-accessible detector that needs no install, no account, and no technical expertise, aimed at the people most targeted (e.g. the elderly).
How It Works
Gemma 4 E4B ships with a dedicated multimodal audio encoder. TrueVoice repurposes it as a feature extractor for anti-spoofing — a task it was never trained for.
- Raw audio → mel spectrogram → frozen Gemma 4 audio tower → 1536-dim representation (mean-pooled)
- A lightweight classification head on top:
Linear(1536→256) → GELU → Dropout → Linear(256→2) - The entire backbone is frozen; only a 1.5MB head is trained. This shows Gemma 4’s pretrained audio representations generalise to security-critical tasks without fine-tuning — and keeps inference fast enough for the web.
The Key Insight: Codec Simulation
Anti-spoofing models trained on studio-quality audio fail on real phone calls, because they learn to judge audio quality rather than genuine synthesis artifacts. TrueVoice applies codec simulation — downsampling training audio to 8kHz and back — forcing the model to learn the acoustic signatures of synthetic speech instead. This was the single most consequential design decision.
Results
Evaluated on the full ASVspoof 2021 LA eval set (148,176 samples):
| Metric | Score |
|---|---|
| EER (Equal Error Rate) | 5.20% |
| Overall accuracy | 96% |
| Fake recall / precision | 96% / 0.99 |
| Real recall | 93% |
Reaching 5.20% EER with a frozen backbone and a 1.5MB head demonstrates the quality of Gemma 4’s audio representations as a general-purpose feature extractor.
The Demo
TrueVoice runs as a Gradio web app in any browser. A user receiving a suspicious call can record it or upload audio and get a verdict in seconds — no install, no account. A “Phone Call Simulation Mode” applies 8kHz codec processing to mimic real phone conditions.
What’s Next
- Telecom API integration for passive call screening at the network level
- On-device inference (Gemma 4 E2B) for fully offline, private detection
- Expansion beyond English — voice phishing is global, and many high-risk groups speak under-represented languages
Tech Stack
- Model: Gemma 4 E4B (
google/gemma-4-e4b-it), frozen audio tower - Training: Google Colab A100, Hugging Face Transformers
- Data: ASVspoof 2019 LA (train) + 2021 LA (eval)
- Demo: Gradio (browser-accessible)
- Classifier size: 1.5MB · Inference: ~2s per clip