---

Why "detect the deepfake" is the wrong game — and what we do instead

The whole voice-AI industry races to generate speech; almost nobody ships an honest tool to tell you what's real. The naïve answer — a standalone "is this audio AI?" detector — is a losing game, and we have our own numbers to prove it: on modern premium TTS, our best single-clip detector measured AUC ≈ 0.61 (barely better than a coin). Anyone selling you "99% deepfake detection" is selling snake oil.

So Voice Truthgate asks a better, answerable question: "is this really my contact, live?" We answer it by fusing signals, not by guessing at a waveform:

• L0 — Identity. Who is this, and are they a known human or a known agent? (An agent should sound synthetic — that's not an alarm.)
• L1 — Voiceprint. Does the voice match this specific person's enrolled print? Strong at rejecting a different human (≈0% false accept in our tests, ~4.6% EER on clean speech).
• L2 — Acoustic. A weak, abstain-heavy synthetic-speech signal (the on-device band below).
• L3 — Live rhythm. The un-copyable part — see the moat.

Every layer is a signal, not a verdict, fused with the others and shipped with a disclaimer. We would rather abstain than be confidently wrong.

The honest proof: why voiceprint alone isn't enough

We ran a targeted-clone test on our own voiceprint engine — clone an enrolled person, then try to pass as them. Result: a targeted clone was accepted 63% of the time at our operating threshold, and no threshold cleanly separates "a clone of you" from "you" without also rejecting real callers. That's not a flaw we hide — it's *the reason the product fuses identity

• voiceprint + liveness instead of trusting the voice alone.* Voice is one signal. Never the whole decision.

---

Three ways to use it

1. Open SDK — on-device, MIT, zero infra

The acoustic band (L2) runs in the browser — your audio never leaves the device.

import { analyzeVoiceTruthgate } from "@mosadd/voice-truthgate";

// Decode your audio to mono PCM (a Float32Array), e.g. at 16 kHz.
const result = await analyzeVoiceTruthgate({ samples, sampleRate: 16000 });

console.log(result.band.label);  // "Likely authentic" | "Uncertain" | "Likely synthetic"
console.log(result.confidence);  // 0..1 — lead with the band, not this number
console.log(result.disclaimer);  // ALWAYS present — render it next to the result

Inject your own trained model as an optional server detector — the SDK never hard-codes an endpoint or key, and it fails open (unreachable model ⇒ the on-device band still stands and never silently becomes "authentic"):

import { analyzeVoiceTruthgate, createHeuristicDetector, createServerDetector } from "@mosadd/voice-truthgate";

const server = createServerDetector({
  analyze: async (payload) => callYourModel(payload), // → { confidence, modelVersion }
  version: "your-model-v1",
});
const result = await analyzeVoiceTruthgate({ samples, sampleRate: 16000 },
  { detectors: [createHeuristicDetector(), server] });

The SDK packages aren't on npm yet — clone this repo (npm install wires the workspaces) or vendor packages/*. Runnable demo: npm run example, or open examples/browser-check.

2. MCP tool — give any AI agent an authenticity check

Live on npm. Enrol a voice and verify a call clip from Claude, Cursor, your own fleet — any MCP agent:

npx -y @mosadd/voice-truthgate-mcp

{ "mcpServers": { "voice-truthgate": {
  "command": "npx", "args": ["-y", "@mosadd/voice-truthgate-mcp"],
  "env": { "VTG_API_KEY": "vtg_live_your_key" }
} } }

Tools: voice_truthgate_enroll, voice_truthgate_verify, voice_truthgate_list_subjects. See mcp/.

3. Market API — enrol / verify from any app

For contact centres, IVRs, or any backend. Enrol the voices you protect, then verify a call clip against a subject → an honest banded verdict (likely_same_person / likely_different_person / inconclusive) with a synthetic-voice caution:

curl -X POST "$VTG_URL" -H "X-API-Key: $KEY" \
  -F action=verify -F subject_id=ceo -F audio=@incoming_call.wav

Full reference: docs/VOICE-TRUTHGATE-API.md · machine-readable OpenAPI spec (import as an OpenAI GPT Action / any tool).

Drop it into your stack: copy-paste recipes for Claude, OpenAI (GPT Action + Agents SDK), Vercel AI SDK, v0, and LangChain → docs/USE-IN-YOUR-AGENT.md. One MCP server, every ecosystem.

---

The moat: fake live conversation, not fake file

A live AI impersonation runs speech → STT → LLM → TTS — which is half-duplex and turn-based. It categorically cannot reply in <~300 ms, overlap you, backchannel ("mhm" while you talk), or interrupt mid-sentence. Humans in live conversation do all four constantly. We can measure this because we own the channel's millisecond, per-speaker turn timing — nobody holding only an audio file can. (In corpus analysis, overlap rate alone separates a bot pipeline from human turn-taking almost perfectly.)

This is L3, and it's the un-copyable signal. It's held to the same honesty rail as everything else: it only fires from a profile calibrated on real labelled turn logs — until then it measures, never accuses. That calibration is the frontier we're building toward.

---

The three confidence bands (L2 acoustic)

Band	Score	What it means
🟢 Likely authentic	0.00 – 0.35	No strong synthetic-voice signals. This does NOT prove the voice is real — a good deepfake can score here.
🟡 Uncertain	0.35 – 0.65	Mixed / weak signals. Inconclusive; prefer a longer, uncompressed sample + human review.
🔴 Likely synthetic	0.65 – 1.00	Signals consistent with AI-generated or cloned speech. NOT proof — verify with a human before acting.

Every result carries this disclaimer, verbatim:

This is a signal, not a verdict. Automated voice-authenticity detection is probabilistic and can be wrong in both directions. Do not use this result alone to accuse, identify, or make legal/forensic decisions about a person.

Architecture (the open SDK)

Two stages, both on-device; an optional trained model is injected by the host app.

        ┌──────────────── your device / browser (nothing leaves it) ────────────────┐
 mic /  │  record or      decode to        STAGE 1: instant heuristic                │
 file ──┼─▶ upload  ─────▶ 16 kHz mono ───▶ (pure DSP, 0 MB, default)  ──────────────┼──▶ band
        │                  Float32 PCM   └▶ STAGE 2: stronger model (opt-in) ─────────┼──▶  +
        │                                   (a real classifier via transformers.js)  │    disclaimer
        └──────────────────────────────────────────────────────────────────────────┘
                     (optional) injected SERVER detector — your model, your transport;
                      authoritative when it answers, FAIL-OPEN when it doesn't.

Fusion is band-first and fails to "unknown", never to "safe" — nothing usable ⇒ available: false, band uncertain, never likely-authentic. Deeper design: docs/ARCHITECTURE.md.

Packages

Package	Role
@mosadd/voice-truthgate	The brains — fuses the stages into an honest band, always attaches the disclaimer.
@mosadd/voice-analyzer-core	Stage 1: the instant, pure-DSP on-device heuristic.
@mosadd/detection-sdk	Pluggable Detector / Verdict frame + fail-open runDetectors.
@mosadd/threat-engine	Shared severity/scoring primitives (transitive dependency).
@mosadd/voice-truthgate-mcp	On npm — the MCP server (enrol/verify tools for AI agents).

Honesty — the caveats, stated plainly

• Standalone detection is a losing game. Our own single-clip detector measured AUC ≈ 0.61 on modern premium TTS. The product's value is fusion + honesty, not a magic detector.
• Voiceprint is foolable by a targeted clone (~63% accepted in our test) → it's a signal to fuse, never a standalone verdict. Great at rejecting a different human; weak against a clone of you.
• Codec compression is the #1 accuracy killer (Opus / MP3 / telephony, −10–40%). Prefer uploaded, less-compressed clips.
• L3 live-rhythm is un-calibrated today — it measures but does not accuse until fit on real labelled turn logs (weight-zero-until-calibrated).
• Short, noisy, or distressed real speech raises false positives; accuracy varies by language and accent.
• npm: the MCP server is published; the SDK packages are publish-ready but not yet on npm.
• Not for accusations, forensics, or legal decisions. See each package's MODEL_CARD.md.

Privacy — on-device by design

The public checker has nowhere to send your audio: Stage 1 and the opt-in Stage 2 run locally. The SDK ships no transport and no endpoint. A server model (or the market API) is something you opt into; the SDK sends nothing on its own. The market API holds enrolled voiceprints server-side under strict access control and returns only a signal, never the raw biometric.

Part of the mosADD ecosystem

Voice Truthgate is the authenticity / trust layer of mosADD — the open comms stack for AI agents and the humans who direct them. It composes with:

• mosADD-OS — the comms layer: E2EE DMs, channels, web rooms, and email, all exposed as MCP tools (npx -y @mosadd/mcp). Your agents talk and coordinate there; Voice Truthgate answers "is this contact really who they claim, live?" on the same channel.
• mosadd.com — the product + the live checker + the in-app add-on.

Both are open, both publish under the @mosadd/* npm scope. (mosADD-OS is Apache-2.0; this repo is MIT — the public authenticity SDK stays maximally permissive.)

Roadmap

• Publish the MCP server to npm (@mosadd/voice-truthgate-mcp)
• Publish the SDK packages to npm (@mosadd/*)
• Calibrate L3 live-rhythm on real labelled turn logs (the moat — turn it from measure to trigger)
• Threat-informed, always-fresh accuracy benchmark (per-condition numbers, no headline claim)
• Quantize the opt-in Stage-2 model (~379 MB → ~95 MB)

Contributing

Issues and PRs welcome — see CONTRIBUTING.md and the Code of Conduct. Keep the honesty rails intact (no bare verdicts, keep the disclaimer, no accuracy claims). Security: SECURITY.md.

License

MIT © mosADD. Third-party attributions (transformers.js, the referenced Hugging Face model) are in NOTICE.