Reachy Mini goes fully local!
Posted by futterneid@reddit | LocalLLaMA | View on Reddit | 8 comments
Hi! Andi from Hugging Face here! My team has been working over the last few months on creating a super smooth local experience for conversations with Reachy Mini, see the video! We hope people can extend this into tons of different cool use-cases.
We wrote a blog explaining how to set this up, and how to modify it for tons of different use cases. Even if you don't have a Reachy Mini, you can use this as a roadmap for amazing voice agents: https://huggingface.co/blog/local-reachy-mini-conversation
Hope you enjoy it!
croholdr@reddit
can you offload some of the models to seperate computers or would that add too much latency?
CryptographerLow6360@reddit
love my mini reachy lite
Fabulous_Fact_606@reddit
Very cool with interruption handling. Now have it hook up to persistent memory context injection...
LKama07@reddit
Great job! Could you please share the delays of each stage of the pipeline?
vulcan4d@reddit
Love it. Tried to do something similar for my kids but quickly realized I had the engineering skills but lacked knowledge getting the models running for some form of real time chat. Nice to see that it is possible.
futterneid@reddit (OP)
You can try it out here: https://github.com/huggingface/speech-to-speech
Even without the robot it's easy to start talking to the console and it feels really cool!
Weary-Step-8818@reddit
fully local is the right default for voice agents. cloud latency makes demos look fine and real interaction feel slightly haunted. the killer test is interruption handling, not just response quality.
futterneid@reddit (OP)
Totally! interruption handling is difficult! I have an approach that I really like here: https://github.com/huggingface/speech-to-speech/pull/255
But every day I think I will merge it and then I keep on finding issues :sob:
Another point for local models with a robot like reachy is that it's constantly running in your home, so privacy can be a real concern