I’d like to share a custom integration that adds Meta’s Muse Voice Transcribe as a native speech-to-text engine for your Assist pipelines.
Why you might like it:
- Truly streaming: audio is forwarded in 80 ms frames while you speak, with interim transcripts arriving mid-utterance. No record-then-upload round trip. In my testing the first partial lands ~0.3 s after audio starts, and the final transcript ~0.4 s after you stop talking.
- Very cheap: Muse bills per second of audio (rate card $0.18/hour ≈ $0.00005/s). A typical 3-second voice command costs a fraction of a cent. To keep it that way, the integration tracks daily usage, supports a configurable per-day limit (it refuses to connect once you’re over and nothing gets billed), and has a
reset_daily_usageservice for one-off overrides. - Keyword biasing: feed it your entity/room names (“Katarina”, “Z-Wave”, “workbench”) and it hears them correctly. 25 languages supported.
Install: HACS → Custom repositories → https://github.com/zraken/ha-muse-stt (category: Integration), then download and restart. You’ll need a Meta Model API key from dev.meta.ai.
Note: it’s a paid API, so setup includes a billing acknowledgement and I’d suggest starting with the daily limit set low.
Happy to hear feedback, especially latency numbers from other setups!