Many voicebots these days need the speed of a fast LLM but the reasoning capabilities of a heavier LLM. To achieve this, modern products are shifting to a dual-model framework. An example is the recent GPT-Live release where a realtime speech-to-speech model has access to GPT5.5 in the background.

This is not a new design, but most of us simply use a heavier model in the background as a tool instead of trying out advanced techniques. For this issue of Computational Conversations I design a few of these techniques to use a fast foreground and slow background model.

I also reproduce a common speech-to-speech pattern of injecting knowledge from a larger LLM live midway (example from Sakana AI) in bot utterance without training a new model, which might be interesting if you like exhausting engineering options before going to model training.

Read more in the full post here.

Feel free to share this email around. Hoping to see you in next month or two!

Abhinav Tushar
https://lepisma.xyz/wiki/about