- Uses Mimi to encode/decode audio tokens, a language model for linguistic generation, and parallel streams for user and assistant.
- Interesting part is not just low latency: listening and speaking are concurrent streams, so overlap and interruption are part of generation rather than external pipeline logic.
- Inner text stream improves linguistic structure before audio-token generation.
Connections
- part_of Speech-to-Speech
- uses Full Duplex
- uses Inner Monologue
- uses Mimi