Realtime voice / live audio agents

When to use

• The user wants a talking agent: speak into a mic, the agent replies with voice, hands-free, low latency (Gemini Live or OpenAI Realtime).
• They want live transcription of the user's speech while the agent talks.
• They want a voice pipeline over an existing text Agent: transcribe a recorded clip → run the agent → (optionally) speak the reply.
• They want to add spoken output (TTS) to an otherwise text-only streaming Agent.

If the user only needs to send a recorded audio file into an agent as one input (not a live session), use ag2-multimodal-input (AudioInput) instead.

Hardware / keys caveat. A real session needs (a) a provider API key and (b) a working microphone + speaker. Neither is available headless. Everything in this skill constructs without hardware; the parts that actually open a socket or audio device are marked [needs keys + audio] below.

Installation

The live module splits across optional extras — install the ones you need: