I built a python script to handle a two-way voice conversation between a user and Anthropic Claude.

[Edit] I just open sourced the code: https://github.com/ccappetta/bidirect...

It uses the open source Faster-Whisper model to transcribe my voice. It has some fairly hairy logic to chunk through the streaming @anthropic-ai Claude replies and send the chunks off to @elevenlabsio for audio generation. Then it's queueing and playing back the audio files with PyGame.

In this way the design is able to generate audio and begin verbalizing Claude's reply within a few seconds of being 'handed the microphone', while the LLM is still continuing to generate that reply in parallel.

Was this the first human-to-AI podcast? I haven't seen any others ¯\_(ツ)_/¯

By the end of the conversation I asked this instance of Claude if it would like to name itself, and it chose "Quill."

I'm also pretty sure Quill was trolling me once I told it that a real-time text to speech design was verbalizing it's replies in the voice of a distinguished British fellow.

All views are my own.