Local AI voice control with Domoticz, n8n, Ollama and faster-whisper - faster than cloud

[xunil74]

I built a fully local AI-powered voice control system using Domoticz + n8n + MQTT.

The system transcribes Hungarian voice commands, queries the Domoticz device list via HTTP API, and uses an AI Agent (Qwen2.5:7b via Ollama) to generate the appropriate JSON control command — delivered via MQTT.

I benchmarked three versions: cloud (Groq + OpenAI), local GTX 1050 Ti, and local RTX 4060 Ti. The local setup ended up 2.4× faster than the cloud.

Full article with architecture, benchmark data and key takeaways:
https://dev.to/xunil74/i-made-local-ai- ... urney-2cko

[gizmocuz]

Great looking project!

I will have a look at this too!

I am using Amazon echo devices to control Domoticz, via a ESP32.

Aldo the switching stays internally, I always thought that it was a 'offline system', until recently the internet was down and I could NOT use any Alexa device.

Did you also investigate a offline speech to text API/processor and just react on words instead of AI?

Like 'Switch on Bedroom Light' it should not be that hard to make a middleware script to understand that a user wants to turn on a bedroom light.
No AI needed for this and much faster too.

[xunil74]

The focus in this project is on real intelligence, not voice control. Voice control is just a sideline, the icing on the cake. As an example, only very simple commands were given, such as "Switch on Bedroom Light", but it can also interpret much more complex commands, just like a human:
- Lower all the blinds in the house except the living room blind
- Lower the blinds in all the rooms in the house to 60%
- I want sunlight in the living room
- Raise the blinds in the rooms where the lights are not on

[gizmocuz]

All seems very well possible bij Speech to Text as well (except the last one, but you can say 'Turn on All lights' of course

You also do not need all of the external applications (single point of failure) and a RTX 4060TI.

Did you had a look Rasspy? Maybe with the help of satellites (ESP32)

https://www.youtube.com/watch?v=gh56Zrw5PI8

You're project is great! But I don't have a RTX 4060 TI in my odroid/raspberry/nuc

[xunil74]

Looks interesting, thanks, I'll check it out.

[mgugu]

I investigated some equivalent function with mcp (latest version of domoticz includes mcp), Claude.AI API, and a single docker container for voice recognition or CLI and exchanges with claude API. Very promising, it is not only possible to send vocal 'intelligent commands to domoticz but also for example to use CLI to add/improve scripts with high level prompts. Main drawback is the price of API tokens, I evaluated around 1€/day for a regular usage. So for me, it is OK for test but not suitable at the moment in production. Other drawback is a high latency. But I imagine same drawback with n8m ?

[xunil74]

In the cloud variation (Groq + OpenAI), the latency is median 4s, in most cases below 6s. It is completely bearable, it does not bother me. The API cost is negligible. Groq STT is free, and I bought a $5 token package with OpenAI. I have run a lot of tests and use them for other things, and so far only $0.07 has been spent. The main reason why I started testing local AI is privacy. I recommend reading my article under the URL given in my first post. There is a lot of data about latency in it.

[gizmocuz]

6s is a lot... 4 is a lot... 2 is even a lot, but... ok.. but still a lot.

I think Rhasspy in combination with satellites (ESP based) is the way to go here (STT and TTS)

https://github.com/Romkabouter/ESP32-Rh ... astLed.hpp

With the hardware described here:

https://smarthomecircle.com/How-I-creat ... -assistant

I ordered some AliExpress parts enough for 3 of them and hope to do some experimentation in a few weeks.

We should leave AI for what it does best... Indeed in combination with MCP, retrieve great values/states from Domoticz, build things (like scripts,automations)
(Claude code ... things like, investigate why my heatpump did not turn on yesterday even when the electricity price indicated a low/cheap price, or build a automation script to turn on my heatpump when the electricity price is cheap, but always make sure the water temperature is at least xx degrees)

For simple Alexa clones STT/TTS is just perfect... I can't imagine waiting for 2 seconds... I don't have that experience with my current Alexa integration... But that one needs the internet... Even that my ESP emulates a local Hue device

[xunil74]

Good job on your project. Let us know when you're done!

[mgugu]

@gizmocuz
I imagine that a lot of code is needed to connect to domoticz and to build the logic of different static vocal commands, which is not different of other integrations. Personally I am in a google environment and I use nodered/smarthome (local command supported) to connect to domoticz.
This is not fundamentally different of what you propose.
But with AI and MCP, you enter in another world with really new possible features for example :
- AI through MCP discovers automatically you environment and understands your logic, so minimal code is needed
- High level commands like "Bedroom is too hot" and AI will automatically adjust your heating/airconditioner system
- Continuous dialog is supported
- Learning can be implemented. Since AI never store data, this has to be stored locally. This can be done of course by AI by selecting pertinent data from last dialog and update a global prompt to be added in the future requests
But I agree, it is the future, today, price and latency are blocking points
Would be interested by the result of your solution when it is up.