nodecode-talk

Talk: speak answers aloud and turn speech into text, through OpenAI's audio API. A Nodecode rewrite of TheSmokeDev/hermes-talk by TheSmokeDev.

The original plugin adds realtime, duplex voice to an agent: you speak, it answers out loud, and it runs its tools without leaving the conversation. A terminal agent has no microphone of its own, so this rewrite keeps the half of that plugin that works from any session — the two OpenAI audio lanes it carries — and exposes them as three tools:

Tool	What it does
`talk_speak`	Say a text aloud: synthesized with OpenAI's speech model and played on this machine.
`talk_transcribe`	Turn an audio file already on this machine into text.
`talk_listen`	Record the microphone for a few seconds and turn what was said into text.
`talk_status`	This cell's own state: the models, the voice, the key's variable and the two audio CLIs. Read-only, no request.

Ask "read me that summary out loud", "what does this recording say?", or "let me just say it" — the /talk command with no arguments records and transcribes, so a spoken question arrives as text the model can act on.

Get it

Tell the agent "add the talk cell", or run nodecode add nc://talk.

Set it up

An OpenAI API key, in the environment variable the section's api_key_env names (OPENAI_API_KEY unless you say otherwise). Nothing else is required: the cell finds a player and a recorder on its own.

"talk": {
  "api_key_env": "OPENAI_API_KEY",
  "voice": "cedar",
  "player": "",          // empty: afplay, paplay, aplay, ffplay or mpv, first found
  "record_command": ""   // empty: arecord, sox or ffmpeg, first found
}

The player and the recorder are CLIs on your machine. record_command is a command line where %f is the file to write and %s the seconds to record, for a recorder this cell does not know.

Use it

"say that out loud" — talk_speak.
"transcribe notes.m4a" — talk_transcribe, on any audio file the API accepts.
/talk hello there says it; /talk alone records and transcribes; /talk status reports the lane.
talk_speak takes speed (0.25 to 4.0) and refuses a voice outside the ten the section offers, before it asks the API.

Audio and text go to OpenAI and nowhere else. Nothing is stored beyond the audio file each talk_speak writes under Nodecode's home (talk/), which the answer names.

What the rewrite leaves out

The original's realtime call itself — one long-lived duplex session over WebRTC, with barge-in, endpointing, live tool calls and the Discord, dashboard and desktop surfaces — is a runtime of its own, not a tool. This cell speaks and transcribes one request at a time. See NOTICE.