Hacker News Reader

Comment by sofixa | Hacker News Reader

I've used Home Assistant Voice (Preview Edition) and ESP32-S3-BOX-3 alongside Home Assistant 's Voice capabilities - by default there are phrases included, and you can add your own, but you have to be precise - "turn off the lights" and "turn off the lamp" are two separate phrases (or a variable one, which can be expanded by a locally (or not) hosted LLM.

Voice detection is pretty good, especially on the Voice, even from across a room. Not 100% (especially with higher pitched voices), but definitely high and usable. The main downside is the fact that you either need to host a local LLM, which isn't cheap, or predefine all phrases you want.

View on HN

I'd really love something that enabled me to use the hardware without Home Assistant. I know it's OSS (basically ESPhome), but I'd like a server I could point my Voice thing to and be able to receive a webhook with the text I just spoke, and have it read the text that my server returns out loud.

I'd implement the actual smarts myself, I just want the STT/TTS interface sorted.

Home Assistant uses the Wyoming protocol to talk to Whisper, Piper, and the various other components of the speech pipeline. Might be something to check out?

https://github.com/rhasspy/wyoming

This subproject looks interesting too

https://github.com/rhasspy/wyoming-handle-external

Yep, I had seen that but I didn't go deeper than that. I'll have another look, thanks!

What has your experience been with Home Assistant voice? I also have one but still use my Alexa’s. I love HA and it runs my house but my initial tests were not great. I tried both OpenAI and HA cloud assistants but I could have spent more time tweaking things and I bet I could have done a lot by writing my own prompt.

It's OK in terms of voice recognition, as in, it almost always work. But integrating with an external LLM for more flexibility in voice commands has been a meh experience and doesn't work perfectly (to be fair, I'm using an external OpenAI compatible API, so it might be something to do with that).

Yeah, I've tried local LLMs, Open AI, and Anthropic. They all sort-of-work and can feel magical but I feel I need to a lot more tweaking/configuring/prompting to improve the experience.

Reply on news.ycombinator.com