The state of voice-AI audio, October 2026
A snapshot as of 2026-10-17: Null Test scores eleven enhancers against a do-nothing control across nineteen conditions, the control sits at 14.32% WER and our v8_2 row at 16.59%, and the market moved towards primary-speaker isolation built into platforms and STT.
Update 20 October 2026: since this snapshot we published a live-conversation benchmark of v8_2 against ai-coustics Quail VF 2.2 S under two recognisers. With faster-whisper base.en it is parity pooled and neither engine beats raw audio; with Gemini Live input transcription v8_2 beats raw and leads VF S pooled. v8_2 leads with a close-mic talker and a loud background, VF S with a distant talker, on pause suppression and on false barge-ins. The snapshot below is unchanged.
As of 2026-10-17, the published Null Test benchmark scores eleven enhancers against a do-nothing control across nineteen conditions; the control sits at 14.32% WER, our v8_2 engine at 16.59%, and nothing on the table beat the control. In the market, the notable change since August is that primary-speaker isolation is no longer sold only as a separate SDK: it shows up inside a platform price (LiveKit), inside an STT option (AssemblyAI) and as a flagship release (Krisp).
This is a snapshot of published state with no new measurements. Benchmark figures come from benchmark.json (generatedAt 2026-08-14); market items come from our internal research compiled 2026-10-15, and every vendor performance number in it is a company claim.
The benchmark, as published
| Backend | Pooled WER (19 conditions) | Shippable |
|---|---|---|
| raw | 14.32% | yes |
| passthrough | 14.32% | yes |
| aic-quail | 15.18% | no |
| focus | 16.59% | yes |
| aic-quail-vf | 17.59% | no |
| gtcrn | 20.14% | yes |
| fastenhancer-s | 20.33% | yes |
| fastenhancer-l | 20.36% | yes |
| fastenhancer-b | 20.64% | yes |
| fastenhancer-t | 21.66% | yes |
| fastenhancer-m | 22.13% | yes |
| dfn3-hq | 30.69% | yes |
| dfn3 | 34.04% | yes |
Pooled word error rate, faster-whisper base.en, ten speakers, 190 reference words per condition. The untouched audio (raw and passthrough, which are the same signal) is at the top because it won. The ai-coustics rows (aic-quail, aic-quail-vf) are a licensed proprietary baseline and not a shipping code path. The row-by-row account is in our row is on the table.
Two earlier results frame it. Our telephony teardown repeated a community figure, that a default resample path discards roughly 16% of caller audio; our control did not support it, and we retracted it in public, leaving the corrected section in place. And the test set is ten speakers: it separates differences of tens of points, not of one or two.
What changed in the market
| Date | Who | What | Source type |
|---|---|---|---|
| 2026-08-12 | Krisp | VIVA 2.5 / Voice Isolation 2.5, marketed as primary-speaker focus; company claims | company blog / CEO post |
| 2026-03-11 | ai-coustics + LiveKit | Voice Focus 2.0 in LiveKit Cloud (current plugin bundles Voice Focus 2.2) | company |
| 2026 (date not found) | AssemblyAI | voice_focus option on streaming STT, $0.10/hr add-on | company docs |
| 2026-08 (Aug) | LiveKit | per-second metering introduced | LiveKit community post |
| 2026-09-15 | Gemini 3.8 Live released | company blog | |
| 2026-10-07 | Krisp | FP16 CPU optimisation post: +75% to +99% concurrent calls on supported CPUs (company claim) | company blog |
All performance numbers attached to these items (for example Krisp's claimed WER reductions) are the vendors' own and unverified by us. Two further data points, both from our research file: voice-AI equity funding was reported at $1.2B across 25 rounds in January to September 2026 against $585M across 34 rounds a year earlier (Tracxn, retrieved October 2026), and Google released Gemini 3.8 Live on 2026-09-15, which matters because a native-audio model has no separate STT whose WER an enhancer could improve.
The consequence we draw is narrow: "speaker isolation" is now a feature name several vendors use, so a claim needs a measurement on a shared set to mean anything. That is the argument for a public control.
What we will measure next
These are plans, not commitments with dates. First, a second recogniser, since the harness supports Deepgram and AssemblyAI and the ordering should hold on more than one. Second, a larger speaker set, because the current one cannot resolve a point or two. Third, false barge-in measured on interruption events, not only on frame-level VAD false alarms as in why turn detection breaks on TV audio. Fourth, CPU cost per stream, since the cost of running a model is becoming part of the competition.
FAQ
Which voice-AI audio enhancer is best? On our benchmark none beat the untouched audio when pooled over 19 conditions; the best enhancers on shippable terms were within a few points of it. Conditions matter more than the pooled number.
What is Null Test? Our open benchmark: eleven enhancers, a do-nothing control, nineteen conditions, published with its configs.
Is the vendor data in this post verified? No. Market items are company claims or independent reports as labelled, compiled 2026-10-15.