This week I added GPT-Live and Gemini Flash 3.8 live support to term-llm.
This is documented at: Live voice · term-llm docs
Voice was always very interesting to me, but also way too flaky. The new breed of voice models, at last, make them a very workable solution.
I enjoy using this a lot, GPT live never misses a word and always delegates correctly to the main session.
The implementation adds 3 providers, Gemini, ChatGPT (for Codex plan) and OpenAI (API access)
I also added an experimental (default off) control lane. This is a bit tricky to get right and introduces some new latency. Will try to refine it over the upcoming weeks. The control lane lets you switch and search for sessions which could allow you to completely control everything using voice.
One snag though is that on the iPhone if you background a browser or lock a phone it tends to tear down a voice connection which kind of forces term-llm to release an app at some point.
I am loving this feature. Be sure to try it out.