Pipecat Labs
Notes from the lab
Models, experiments and write-ups from the team building Pipecat: open weights for voice agents, the benchmarks behind them, and what we learned along the way.
All posts
02 posts
Research
Experimenting with tiny classifiers for tone and intent
Two tiny CPU classifiers that read how a caller meant each turn and choose how the agent should say its reply, so any LLM in a voice pipeline can sound like it’s listening.
Release
Introducing PhoneLLM Alpha 2
Two open-weights text models for voice agent pipelines — Mini for speed, Max for accuracy — with microthinking, Pipecat Flows support, and more flexible prompting.