#Hardware requirements for Voice Assistant
1 messages · Page 1 of 1 (latest)
What language would you like to use to communicate with the assistant?
Do you want to run TTS/STT/AI services locally, or is the cloud fine for you?
There's no hats for AI, if you plan using local LLM...
Ok, that's good to know
My idea was to run Assist locally (also because I am not always connected to the internet). The language is Dutch, and if that's not developed enough I will use English.
The price difference between the RPI's (4 and 5) is not that big. So I guess if it is a RPi I should go for the 5. I probably should also go for 8GB. And then maybe a SSD or NVMe is needed.
Or perhaps I shouldn't look at RPi's in the first place if that's just a bit too slow. But it's all just guessing for me.
The Dutch intents built into Home Assistant are 100% complete, meaning everything available in English is also accessible in Dutch. You probably won’t have any issues with local TTS in Dutch either. I can’t comment on STT, but as the saying goes, “the proof of the pudding is in the eating.” (In Hungarian, only the large model works, which requires a GPU unless you’re okay with waiting 30-40 seconds.)
However, AI is out of the question. For a normal processing speed, you need a GPU or a very powerful CPU and a lot of RAM. I gave up on it, especially since it’s far inferior to GPT-40-mini, which is both much cheaper and highly efficient (the electricity costs alone wouldn’t fit into the <$1 monthly budget).
Thank you for this clarification.
Ok, so let's forget the AI. Maybe in the future.
What would you recommend if I would only use TTS and what if I also want STT?
I recommend the Piper/Whisper combo locally. It even runs smoothly on a Raspberry Pi in English. If you’re considering a Pi, make sure to get the latest model, especially if the price difference isn’t significant.
If a cloud solution works for you, I personally use Azure TTS/STT and GPT-40-mini.
I also spent a lot of time figuring out what to buy, but in the end, I found refurbished Dell OptiPlex 5070 Tiny PCs and decided to go with those.
Thank you.
What amount of memory do you use in your PC? And do you think I need a SSD/NVMe?
Having SD card for HA is already deprecated approach...
Ok, that's a good statement. Is SSD acceptable, or would you recommend to go for a NVme?
SSD is fine. Don't think it will be bottleneck on Pi5
You can easily run an AI on a good smartphone lately. An 8GB pi5 should handle it fine.
Hmm... https://llamaimodel.com/requirements/
I played around with it quite a bit, but it requires a beast of a GPU. A phone can easily send the query to the cloud via an API 🙂
I wouldn't use the cloud personally, but yes, easier to send to a beefy PC server.
Well, yes, but models without a GPU are heavily limited. On my language, only the large models work at all, and even with a GTX 4070, they're still slow 😦 I don’t have any other alternatives.
Meta have a small & highly multilingual model. They wrote a paper on it.
Yes, it's small and multilingual, but it needs a GPU 🙂 I played around quite a bit with Meta's Llama model.
On a Raspberry Pi, local Hungarian speech recognition is like a dream. Even major phone manufacturers haven’t been able to solve this without an internet connection.
If your goal is simple chat - probably (though on my i7 I have like 8 tokens/sec on llama3.2, which might be ok for chat, but hardly acceptable for voice response, when I have to wait full response generation before TTS). But add tools and context - and it's unbearable.
I just get 8.47 tps on an S22 ultra. llama 3.1 3b q4_k, with the PocketPal app.
Right, but 3b quantified to q4 won't give you good results with smart home.
Also i bet you used it in chat window. Connect it to HA, give "Assist" possibility to it and you will see.