# ai-coustics' Quail Voice Focus 2.3 Cuts Transcription Errors by 11 Points

> Source: <https://alphasignal.ai/news/ai-coustics-quail-voice-focus-2-3-cuts-transcription-errors-by-11-points>
> Published: 2026-11-18 15:31:11+00:00

**Key takeaway:** Upgrading to ai-coustics SDK 0.26.0 requires swapping pinned model keys or cached files to quail-vf-2.3-l-16khz or quail-vf-2.3-s-16khz, since the SDK cannot load Voice Focus 2.2 models.

# ai-coustics' Quail Voice Focus 2.3 Cuts Transcription Errors by 11 Points

ai-coustics' Quail Voice Focus 2.3 keeps the main speaker even when they're quiet or far from the mic. It runs on CPU in about 30 ms and cuts word errors by up to 43%.

- ai-coustics' [Quail Voice Focus](https://huggingface.co/ai-coustics/quail-voice-focus?utm_source=alphasignal) isolates the main speaker for voice agents, on CPU, in about 30 ms.
- Version 2.3 keeps quiet or distant main speakers; SDK 0.26.0 can no longer load 2.2 models.
- Cuts word error rates by 10-43% across seven speech-to-text providers, mostly by removing background speech from transcripts.
- Trained on semi-synthetic audio of real failure cases; tested on the open [Dawn Chorus](https://huggingface.co/datasets/ai-coustics/dawn_chorus_en?utm_source=alphasignal) dataset.
- Single speaker only; needs some audio before suppression fully kicks in, and the focus can shift to a louder speaker.
- Free 30-day trial; [paid plans](https://ai-coustics.com/pricing?utm_source=alphasignal) start at $135/month for 100,000 minutes.

Berlin-based ai-coustics has released version 2.3 of [Quail Voice Focus](https://huggingface.co/ai-coustics/quail-voice-focus?utm_source=alphasignal), a real-time model that removes competing speech before voice-agent audio reaches transcription. The update improves retention of quiet or distant foreground speakers, which earlier versions could partly suppress. The model’s Hugging Face page recorded more than 900,000 downloads in the previous month, although that count reflects file pulls rather than unique deployments.

Background speech creates a costly failure mode for voice agents. A speech-to-text service may transcribe dialogue from a television, coworker, or taxi passenger as part of the user’s request. The language model then receives a plausible transcript with no reliable way to determine which words came from the intended speaker.

Quail Voice Focus sits between audio capture and speech-to-text, or STT. It selects one foreground speaker while suppressing noise, room reverberation, echo, codec artifacts, and other voices. ai-coustics reports 30 milliseconds of processing latency on a CPU, with no GPU or ONNX runtime required. Capture buffers, network transport, and transcription add separate latency to the full pipeline.

## Version 2.3 changes the upgrade path

Version 2.3 ships with ai-coustics SDK 0.26.0 and replaces the 2.2 model files. Developers upgrading the SDK must also update any pinned model key or cached model file.

- **Large model:**`quail-vf-2.3-l-16khz`
- **Small model:**`quail-vf-2.3-s-16khz`
- **Breaking change:** SDK 0.26.0 cannot load Voice Focus 2.2.

According to ai-coustics, the revised model retains the dominant speaker more reliably when that person is quiet, far from the microphone, or using speakerphone. Earlier releases were tuned more heavily for close-talking speech, increasing the risk that a distant caller would be suppressed with the background audio.

## Generic cleanup can preserve the wrong voice

Conventional noise suppression often treats intelligible speech as useful audio, regardless of who is speaking. That behavior works for calls where every participant should remain audible. A single-user voice agent needs one speaker to survive the filter while nearby conversations disappear.

## This story is for Pro members

You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
