@mosadd/voice-analyzer-core

Pure-TS deterministic voice-deepfake feature extractor + scorer. No DOM, no Web Audio, no FFT library — runs identically in browser, React Native, Deno edge functions and Node tests.

License: MIT.

Why this exists

mosadd needs the same scoring function in three different runtime environments without divergence:

1. apps/web Demo112 voice analyzer — minister demo, calls analyzeAudio(samples, sampleRate) on Web Audio captures.
2. supabase/functions/analyze-call-audio — LiveKit webhook handler, decodes egress recordings and emits session_events with the same shape.
3. apps/gov-edge Rust appliance (future) — implements the same weighting in Rust; CI runs the parity fixtures from this package.

If the scoring drifts between consumer banner and operator console, an operator could see "0.91 synthetic" while the user sees "0.06 fine". That is the bug we are preventing.

API

import { analyzeAudio } from "@mosadd/voice-analyzer-core";

const result = analyzeAudio(pcmSamples, 16000);
// {
//   confidence: 0.87,
//   verdict: "VOICE_DEEPFAKE_HIGH_CONFIDENCE",
//   features: { spectralTiltStdDev: 0.05, zcrJitter: 0.002, ... },
//   topReasons: ["heuristic_score=0.87", "prosody_flatness=0.92", ...],
//   analyzerVersion: "heuristic-1.0"
// }

Scoring heuristic

The MVP scorer (version heuristic-1.0) weights five cues:

| Cue | Weight | What it measures | |-----|--------|------------------| | prosodyFlatness | 0.28 | Flat energy envelope = unnatural | | zcrLowJitter | 0.22 | Synthetic voices have unnaturally smooth ZCR | | spectralTiltLowVariance | 0.18 | Real speech tilt wanders, synth holds steady | | noBreathPauses | 0.18 | Real voice has breaths between phrases; synth doesn't | | f0LowVariance | 0.14 | Monotone pitch curve |

All five cues are deterministic time-domain or narrow-band-Goertzel measures — no FFT library, no model file.

Faza B — replacing with AASIST ONNX

When LINEAR-2009 Faza B lands a real AASIST-PL model, this package keeps the same exports but scoreFeatures becomes a thin wrapper around ONNX inference. The wire shape (VoiceAnalysisResult) does not change, so neither the operator console nor the consumer banner needs changing.

Synthesizers

For offline reproducible demos and tests we ship three signal synthesizers (NOT real recordings):

• synthesizeHumanPolish(seed) — varied F0, natural ZCR jitter, embedded breath pauses, dynamic envelope.
• synthesizeElevenLabsLikeSynth(seed) — uniform F0, near-zero ZCR jitter, zero breaths, flat envelope.
• synthesizeVoiceCloneChild(seed) — child-pitch, synthetically smooth.

Real recordings can later land as apps/web/public/audio/*.wav and be decoded by the AudioContext to feed analyzeAudio() — the function itself does not care about the audio source.