here's how it performs as the llm in a voice agent stack. https://github.com/tmshapland/talk_to_gpt_oss