Tune Sentence Breaks and Latency
Overview
KizunaAI translates a sentence once it decides you have finished it. If translations arrive too late, or sentences are cut in the middle, Endpoint Detection Tuning changes how that decision is made. The defaults suit most conversations; change one setting at a time and try it in a short session.
1Open the tuning settings
Open Settings, switch to Advanced with the button at the top right of the panel, and choose the Provider tab. Scroll to Endpoint Detection Tuning. Changes apply from the next session.
2Endpoint Sensitivity
How readily a pause counts as the end of a sentence, from −1 to 1 (default 0). Higher values end sentences sooner, so translations arrive faster but sentences may be cut off early. Lower values wait for stronger evidence that you have finished.
3Max Pause Before Finalizing
A hard limit, from 500 to 3000 ms (default 2000 ms): once you stop speaking, the sentence is closed after at most this long. Raise it if you often pause mid-sentence and your sentences get split; lower it for faster translations.
4Latency Reduction Level
From 0 — Default through Lower latency and Even lower latency to Most aggressive. Each step closes sentences more eagerly to cut delay.
Tips
- Translations arrive late: try raising Endpoint Sensitivity a little, or a higher Latency Reduction Level.
- Sentences are cut in the middle: lower Endpoint Sensitivity, or raise Max Pause Before Finalizing.
- Pausing briefly at the end of each sentence helps whatever the settings.
Frequently Asked Questions
How do I go back to the defaults?
Set Endpoint Sensitivity to 0.0, Max Pause Before Finalizing to 2000 ms and Latency Reduction Level to 0 — Default.
Is there a push-to-talk mode?
Not with KizunaAI. It detects the end of each sentence by itself, which is what these settings tune.
Troubleshooting
Still stuck? Send us feedback from the account menu, or write to support@kizuna.ai. You can also ask in GitHub Discussions.