Is the 0.961 cross-encoder verified?

The headline is encoder-matched (guard and evaluator both WavLM-base-plus-sv). Cross-encoder triangulation with ECAPA-TDNN is instrument-ready and pending. A subsequent version of this manuscript will promote the figure if ECAPA mean agreement lands within ±0.10 and Pearson r ≥ 0.5.

Multi-speaker? Multi-language?

The measured anchor is one speaker, English, encoder-matched. Multi-speaker replication is named the highest-priority next measurement; cross-language pilots are a future direction, not present evidence.

Yes. The reference clips, computed centroid, and pipeline never leave the customer environment.

Best-of-12 is inference-intensive; the reference build runs 10-sentence evaluation in ~9 minutes on a single RTX 4090. Low-latency real-time deployment requires K reduction or distilled scorer — named engineering follow-on.

How is this different from VALL-E 2?

VALL-E 2 claimed human parity on LibriSpeech continuation at WavLM ~0.881 under single-prompt in-context learning. We measure 0.961 on a subject-specific 10-sentence held-out set under 50-clip centroid + best-of-12 selection. Scope is one subject, English; VALL-E 2's scope is LibriSpeech-derivative. Not a benchmark contest.

HOME · APPS · APP 6

APP 6 · VOICE

0.961 mean WavLM SECS. DNSMOS P.808 parity with real recordings. Unmodified Qwen3-TTS + 50-clip centroid + best-of-12 at T=0.3.

Measured 0.961 mean WavLM SECS + DNSMOS 3.93/3.93 parity + per-sentence cosine audit on an unmodified Qwen3-TTS base — the empirical anchor that ports into any frontier-lab TTS surface. ElevenLabs $11B at $330M ARR sits inside a 12–24 month absorption window and the regulated half of the $47.5B voice-AI market — financial-services outbound, healthcare agent-of-record, insurance underwriting — is structurally inaccessible to unverified voice; this is the pattern that opens it. Every output verified against a frozen speaker-embedding centroid, +0.080 over VALL-E 2 human parity, zero human annotators.

MEASURED · 2026-04-16 · N=10 sentences, one speaker, encoder-matched WavLM-base-plus-sv

Schedule 48-hr POC Book 20-min briefing Strategic inquiry

WHO THIS IS FOR

For enterprise brand-voice / executive-comms teams, contact-centre BPOs seeking 30% automation without compliance liability, and verification-grade synthetic-media authentication vendors.

PROOF STACK

Every claim tagged.

MEAN WAVLM SECS

0.961 (N=10, encoder-matched, 2026-04-16)

MAX / MIN

0.975 / 0.914

SENTENCES ≥ 0.95

9 of 10

VS VALL-E 2 HUMAN PARITY

+0.080

DNSMOS P.808

3.93 clone / 3.93 real

UTMOS

3.012 clone / 2.997 real

BOOTSTRAP 95% CI (SEED 20260416, N=10K RESAMPLES)

[0.951, 0.971]

TECHNICAL DEPTH

How it composes.

Base: Qwen3-TTS, inference-only (no parameter update)
Full In-Context Learning over all 50 reference clips
50-clip L2-normalised WavLM-base-plus-sv centroid
Best-of-12 at T=0.3, argmax by cosine against centroid
Threshold τ=0.95 calibrated to same-session-band floor pre-deployment

MARKETS UNLOCKED

What opens downstream.

$47.5B by 2034

Voice AI at compliance grade

$5–20B ARR

Enterprise brand-voice products

$10–40B inside $58.3B fraud pool

Verification-grade synthetic media / anti-deepfake

$10–30B/yr globally

Contact-centre / BPO

FREQUENTLY ASKED

Questions from every buyer.

0.961 mean WavLM SECS. DNSMOS P.808 parity with real recordings. Unmodified Qwen3-TTS + 50-clip centroid + best-of-12 at T=0.3.

MEASURED · 2026-04-16 · N=10 sentences, one speaker, encoder-matched WavLM-base-plus-sv

Questions from every buyer.