Quail Voice Focus 2.1: New flagship speaker isolation, 10x more efficient

Home

Blog

Home

Blog

Quail Voice Focus 2.1: New flagship speaker isolation, 10x more efficient

Written by

Fabian Seipel

CEO and Co-founder

Product

Apr 23, 2026

When deploying voice agents at scale, performance and efficiency are key for any organization. It's the difference between a pipeline that holds up under pressure and one that degrades every time conditions aren't perfect.

Clean audio input is where reliable voice agents start. That’s why we’re launching Quail Voice Focus 2.1, our flagship speaker isolation model rebuilt to be up to 10 times more efficient, while offering better performance than its predecessor.

Why speaker isolation matters

Production voice pipelines have a persistent problem. Without isolation, background speakers and device echo are all treated as intelligible speech. It gets transcribed alongside the primary speaker and corrupts the turn-taking logic, often breaking the agent stack entirely.

Our Quail Voice Focus model is built to fix this at the source by isolating the primary speaker in real-time. As soon as the model recognizes the primary speaker it locks onto the foreground speaker consistently regardless of input loudness, making it one less variable your pipeline needs to account for.

Two new models, built for scale

Today we are releasing Voice Focus 2.1, delivering improved primary speaker isolation at a fraction of the compute:

Voice Focus 2.1 S is 10x smaller than Voice Focus 2.0 and still outperforms it. Built for high call volumes on constrained infrastructure and edge deployments. Choose S for maximum efficiency with a minimal drop in performance.
Voice Focus 2.1 L delivers the best-in-class enhancement quality at 25% lower compute than 2.0. Choose L for best overall performance.

Both models run in real time on CPU with an end-to-end latency of 30ms, work at 8 kHz as well as 16 kHz and drop straight into your existing integration.

Improved real-world performance

Real-world audio is messier than most benchmarks capture. Production calls arrive with multiple simultaneous background speakers, media playback, and layered noise, alongside primary speakers across a wide range of languages, accents and loudness levels. The Voice Focus 2.1 training pipeline was built around that reality, with significantly more complex acoustic scenes than previous generations pushing robustness further than ever before.

Quail Voice Focus 2.1 Benchmark

Bar charts comparing Word Error Rates on commercial STT models across unenhanced audio and audio enhanced by ai-coustics VF 2.0, VF 2.1 S, and VF 2.1 L. The graphs show ai-coustics VF 2.1 L achieves the lowest total WER across all models, primarily by reducing substitution errors. Bars exceeding the 32% chart range are truncated; the true total WER is noted. Full results: For AssemblyAI Universal-3 Live, deletions are 4.7% for unenhanced, 8.3% for VF 2.0, 9.0% for VF 2.1 S and 8.4% for VF 2.1 L; insertions are 7.9% for unenhanced, 4.2% for VF 2.0, 4.8% for VF 2.1 S and 4.3% for VF 2.1 L; substitutions are 35.7% for unenhanced, 6.3% for VF 2.0, 3.7% for VF 2.1 S and 2.4% for VF 2.1 L; total WER is 48.3% for unenhanced, 18.8% for VF 2.0, 17.5% for VF 2.1 S and 15.1% for VF 2.1 L. For Deepgram Nova 3 Live, deletions are 2.1% for unenhanced, 3.7% for VF 2.0, 3.9% for VF 2.1 S and 3.2% for VF 2.1 L; insertions are 6.0% for unenhanced, 4.4% for VF 2.0, 4.7% for VF 2.1 S and 4.4% for VF 2.1 L; substitutions are 64.2% for unenhanced, 14.6% for VF 2.0, 8.1% for VF 2.1 S and 7.9% for VF 2.1 L; total WER is 72.3% for unenhanced, 22.8% for VF 2.0, 16.6% for VF 2.1 S and 15.6% for VF 2.1 L. For Soniox STT Async v4 Live, deletions are 5.4% for unenhanced, 3.6% for VF 2.0, 2.9% for VF 2.1 S and 2.2% for VF 2.1 L; insertions are 4.0% for unenhanced, 3.3% for VF 2.0, 3.5% for VF 2.1 S and 3.5% for VF 2.1 L; substitutions are 67.7% for unenhanced, 11.5% for VF 2.0, 9.4% for VF 2.1 S and 7.4% for VF 2.1 L; total WER is 77.1% for unenhanced, 18.4% for VF 2.0, 15.9% for VF 2.1 S and 13.2% for VF 2.1 L. For Mistral Voxtral Mini Live, deletions are 6.6% for unenhanced, 7.9% for VF 2.0, 6.6% for VF 2.1 S and 5.1% for VF 2.1 L; insertions are 3.3% for unenhanced, 3.6% for VF 2.0, 3.5% for VF 2.1 S and 3.6% for VF 2.1 L; substitutions are 29.8% for unenhanced, 5.7% for VF 2.0, 4.6% for VF 2.1 S and 3.4% for VF 2.1 L; total WER is 39.7% for unenhanced, 17.1% for VF 2.0, 14.7% for VF 2.1 S and 12.1% for VF 2.1 L. For Cartesia Ink-Whisper Live, deletions are 3.0% for unenhanced, 5.5% for VF 2.0, 5.3% for VF 2.1 S and 5.2% for VF 2.1 L; insertions are 7.2% for unenhanced, 5.8% for VF 2.0, 6.8% for VF 2.1 S and 6.7% for VF 2.1 L; substitutions are 42.7% for unenhanced, 8.6% for VF 2.0, 5.7% for VF 2.1 S and 4.8% for VF 2.1 L; total WER is 52.9% for unenhanced, 20.0% for VF 2.0, 17.8% for VF 2.1 S and 16.8% for VF 2.1 L. For Gladia Live, deletions are 4.0% for unenhanced, 3.4% for VF 2.0, 3.9% for VF 2.1 S and 3.9% for VF 2.1 L; insertions are 4.7% for unenhanced, 4.6% for VF 2.0, 5.4% for VF 2.1 S and 5.3% for VF 2.1 L; substitutions are 16.6% for unenhanced, 10.6% for VF 2.0, 5.9% for VF 2.1 S and 5.0% for VF 2.1 L; total WER is 25.2% for unenhanced, 18.7% for VF 2.0, 15.3% for VF 2.1 S and 14.2% for VF 2.1 L. For Speechmatics Live, deletions are <1% for unenhanced, 2.8% for VF 2.0, 2.8% for VF 2.1 S and 3.1% for VF 2.1 L; insertions are 4.4% for unenhanced, 3.2% for VF 2.0, 3.1% for VF 2.1 S and 3.2% for VF 2.1 L; substitutions are 60.5% for unenhanced, 11.4% for VF 2.0, 7.5% for VF 2.1 S and 6.2% for VF 2.1 L; total WER is 65.8% for unenhanced, 17.3% for VF 2.0, 13.4% for VF 2.1 S and 12.5% for VF 2.1 L. For Gradium Live, deletions are 2.9% for unenhanced, 4.8% for VF 2.0, 5.3% for VF 2.1 S and 5.7% for VF 2.1 L; insertions are 6.9% for unenhanced, 6.8% for VF 2.0, 6.9% for VF 2.1 S and 7.1% for VF 2.1 L; substitutions are 40.6% for unenhanced, 10.6% for VF 2.0, 6.9% for VF 2.1 S and 5.1% for VF 2.1 L; total WER is 50.4% for unenhanced, 22.2% for VF 2.0, 19.1% for VF 2.1 S and 17.8% for VF 2.1 L.

We evaluated Voice Focus 2.1 on a data set that mimics real-world failure cases closely, see this page for a range of qualitative examples.

Processing raw audio leads to a high volume of insertions across all major STT providers, resulting in broken turn-taking and polluted LLM context. Across seven major STT providers, unenhanced audio sits between 25–77% WER.

Voice Focus 2.1 cuts that dramatically, on Speechmatics from 65.8% down to 12.5%, on Cartesia from 52.9% to 16.8%, on AssemblyAI from 48.3% to 15.1%. The gains are consistent across every provider tested, driven by a sharp reduction in insertions with only a marginal increase in deletions.

Voice Focus 2.1 S and L deliver up to 10x more concurrent streams at 30% safety margin, tested across three major AWS instance types. As the graph below shows, both models outperform Voice Focus 2.0 while running significantly leaner on CPU.

Bar chart showing max concurrent streams by model and instance type, comparing ai-coustics VF 2.0 (baseline), VF 2.1 L, and VF 2.1 S across three AWS 8xlarge instances. VF 2.1 S achieves the highest throughput on all instance types, reaching the 180-stream benchmark cap on AMD and Graviton. Safety margin is 30%; benchmark is capped at 180, and top values may exceed this. Full results: For c7a.8xlarge (AMD EPYC 3.7GHz), max concurrent streams are 46 for VF 2.0, 130 for VF 2.1 L, and 180+ for VF 2.1 S. For c7g.8xlarge (AWS Graviton 2.6GHz), max concurrent streams are 40 for VF 2.0, 71 for VF 2.1 L, and 180+ for VF 2.1 S. For m7i.8xlarge (Intel Xeon 3.2GHz), max concurrent streams are 32 for VF 2.0, 82 for VF 2.1 L, and 132 for VF 2.1 S.

The result is lower Word Error Rates and stronger VAD performance on the audio that actually shows up in production, for both the built-in ai-coustics VAD and pipelines pairing Voice Focus with external VADs like Silero.