Whisper Name Biasing

Published 2026-09-09

As of version 3.7.0, the whisper app has gained the ability to bias speech-to-text towards your Home Assistant entity, area, and floor names. This means that names like "EcoBee" will be correctly recognized instead of "he go be" or something else. Enable the bias_names option and restart the app to try it out.

Name biasing injects area, floor, and exposed entity names into the prompt of capable speech-to-text libraries like faster-whisper, transformers, and the new qwen3-asr (slower but more accurate than auto). With the names in the prompt, these libraries are better able to resolve unusual names correctly. There is limited space, however, so names are prioritized:

1. Areas and floors with exposed entities in them
2. Exposed entities from high priority domains, such as lights and scenes
3. All other area/floor names
4. All other exposed entity names

Any name that falls outside the maximum prompt size is dropped. If you want more control, you can set the initial_prompt option manually to include specific vocabulary. If bias_names is enabled, your manual prompt is included first.