Project 06 · Audio Forensics & Security

VoiceGuard

Continuous acoustic verification and deepfake synthesis detection for live calls, streams, and virtual conferences — using a fine-tuned self-supervised speech encoder, entirely on-device.

Overview

VoiceGuard detects deepfake and synthetic voice attacks in real time, across both live voice streams and recorded audio files. It defends voice channels against impersonation and against synthetic voice generators such as ElevenLabs, ChatTTS and Edge-TTS.

Rather than returning an opaque verdict, it produces a transparent 0–100 Live Risk Score built from three independent forensic signals — so a reviewer can see why a sample was flagged, not just that it was.

At a Glance

~13 ms

Inference per 4s window

On GPU

<200 ms

Inference per 4s window

On CPU

98.9%

Unseen TTS detection

Held-out generators

16 kHz

Live PCM stream, mono

Sliding 4s analysis

100%

Local, on-device inference

No persistent storage

Tri-Signal Architecture

Three independent forensic signals are computed in parallel, then fused into one explainable score.

Raw Audio Stream / File ↓ Preprocessing (16 kHz mono) ↓ ┌────────────────┼────────────────┐ ↓ ↓ ↓ wav2vec2 ECAPA-TDNN Praat / Parselmouth (fine-tuned) (SpeechBrain) Prosody forensics Synthetic Speaker Naturalness speech detect verification F0 · jitter · shimmer · HNR │ │ │ Bonafide (0-1) Cosine sim (0-1) Naturalness (0-1) └────────────────┼────────────────┘ ↓ Explainable Fusion Scorer ↓ Risk Score (0-100) ↓ Live Risk Dashboard & Alerts

Fusion Engine

The three signals combine through transparent, configurable weights. No hidden ensemble — the contribution of each signal is inspectable and tunable at runtime via the config endpoint.

Risk = ( w_spoof · (1 − S_bonafide) + w_speaker · (1 − S_speaker) + w_prosody · (1 − S_naturalness) ) × 100

Key Features

Synthetic Voice Detection: a fine-tuned wav2vec2 model (with AASIST-L support) trained across multiple TTS systems to catch synthesis artifacts.
Speaker Verification: ECAPA-TDNN 192-dimensional embeddings compare the incoming voiceprint against an enrolled identity profile.
Clinical Prosody Forensics: Praat/Parselmouth measures acoustic micro-variation — F0 pitch variability, jitter, shimmer, and harmonics-to-noise ratio.
Explainable Weighted Fusion: the three signals combine through transparent, configurable weights rather than an opaque ensemble.
Real-Time Streaming: live 16 kHz PCM over WebSockets with a rolling buffer — roughly 13 ms inference per 4-second window on GPU, sub-200 ms on CPU.
Privacy by Design: 100% local, on-device inference. Audio is processed in memory and discarded immediately, with no persistent storage.

API Surface

Analysis

POST /analyze WS /ws/analyze

Identity

POST /enroll

Configuration

/config

Tech Stack

Backend

Python 3.10+ PyTorch FastAPI WebSockets

Audio & Models

wav2vec2 AASIST-L ECAPA-TDNN SpeechBrain Librosa Praat / Parselmouth

Frontend

Next.js 16 React TypeScript Tailwind CSS Lucide Icons

Interface

Overview

Headline detection claim, live capability figures, and entry points into the detector and docs.

Live Detector

Real-time waveform with the verdict, a 0–100 risk score, and a confidence band as audio streams in.