> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ai-coustics.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Voice Focus for voice AI systems

> Tune primary-speaker isolation against transcription errors and speech retention.

**Quail Voice Focus** is designed for voice agents that should follow one primary speaker. Evaluate it with your speech-to-text (STT) provider to determine whether suppressing competing speech improves the behavior your application needs.

Use the [paired audio evaluation procedure](/guides/evaluate-audio) to establish a baseline, then rerun the same recordings after each tuning change.

`ProcessorParameter.EnhancementLevel` accepts values from `0.0` to `1.0`. Higher values request stronger suppression. The starting value comes from the loaded model; read it through `processor.get_context().get_parameter(...)` in Python or `processor.getContext().getParameter(...)` in Node.js.

<AccordionGroup>
  <Accordion title="How it works">
    Voice Focus makes suppression decisions from audio context. Speech can be ambiguous, especially when speakers overlap or the target is quiet. Increasing the enhancement level changes suppression strength; it does not guarantee that every foreground word will remain or that every background voice will disappear.
  </Accordion>
</AccordionGroup>

## Recommended settings

Use these as comparison points for an evaluation, not measured quality rankings:

| Value | Evaluation purpose | Check |
| - | - | - |
| Current model default | Establish the baseline for the exact model artifact. | Speech retention, background leakage and startup behavior. |
| `0.5` | Compare a lower setting with stronger suppression. | Whether quiet target speech is retained and whether competing speech reaches STT. |
| `0.8` | Test increased suppression when competing speech causes errors. | Whether fewer background insertions justify any added target deletions. |
| `1.0` | Test the strongest exposed setting. | Quiet speech, overlaps and speaker changes; do not assume this minimizes word error rate. |

Keep the input recordings, STT version and decoding settings fixed. Change one setting at a time. No setting guarantees preservation of all foreground speech, and the best result can differ by provider and acoustic condition.

## Best practices

* **Match the target-speaker task.** Use Voice Focus when background speech should be suppressed. Use [Quail's multi-speaker models](/models/speech-enhancement/quail-multi-speaker) when all participants matter.
* **Measure insertions and deletions separately.** Lower overall word error rate can hide loss of a critical target utterance.
* **Include startup and retargeting.** Test the beginning of a stream, quiet speech and a change in the most prominent speaker. Preserve state between consecutive frames.
* **Check agent behavior.** Compare false interruptions, missed requests and end-of-turn latency alongside transcription quality.

Use the [model reference](/reference/sdk/models#quail-voice-focus) to record the exact model ID and the [performance guide](/reference/concepts/performance) for reproducible resource measurements.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.