Skip to content

insodimension/voice-mode

v0.1.3MIT

Voice mode for any agent in any space: the Voice pane in Settings shows which speech engines are ready, which voices exist (yours, your workspace's, a plugin's) and which agent sounds like what, and which live voice Talk live would open. The speaking itself is the engine's reply vocalizer; open voice options from the small dot left of the composer microphone to turn it on or off. The microphone itself is for dictation.

GitHub

Voice mode

What it is

The Voice pane in Settings: it tells you whether voice mode can speak right now, which speech engines are ready, which voices exist and what each would really speak with, which agent sounds like what, and how to change any of it. Voice mode itself (the engine's reply vocalizer, the voice desk, the composer's speaker toggle) is the engine's, not this pack's; this pack is the window onto it.

What it contains

One component, voice-settings, filling the settings slot with the label Voice. It shows:

  • A headline saying whether voice mode can speak, from the speech profiles the engine publishes.
  • Speech engines: each installed provider, whether it is ready to speak, to listen and to talk live, and why not.
  • Voices: every voice profile visible from the workspace (workspace, yours, a plugin's, built-in), its fallback chain step by step, and the voice it would speak with now; and, when the voice names live choices, the live voice Talk live would open (the first ready one, with no on-device fallback) and why any before it was passed over.
  • A Talk live line under the headline: which live voice the default voice opens, or that it names none. It says nothing on an engine with no live voice at all.
  • Agents: the voice each agent ships with, and the default for the agent that names none.
  • Models: the two small models that make voice mode feel attentive (the voice role that rewrites a reply into spoken prose, and the classifier role). Neither is required. On an engine that says where the classifier role resolves, its row names the model and states what it is sent and where (below); on an older engine it lists the classify models connected and claims nothing more.
  • Change it: where to assign an agent a voice (voice: in agent.md, or voice.agents.<agent> in config.json), where to drop a new voice profile (voice-profiles/), and the voice block keys that tune behaviour (vocalizer.mode, vocalizer.enhanced, attention.catchUpAfterMinutes, attention.chimes, live.idleMinutes), and the converse: list that chooses who talks live.

No tools, skills, prompts or rules. The pack imports only react.

Who can use it

Anyone on the canary release ring (channel: "canary"). It is on by default (defaultEnabled: true) and is listed by the Chat pack, so enabling Chat brings it along. It works in every space and with every agent. No sign-in, key or CLI.

Limits and risks

  • What the classifier is sent. The classifier reads your mood and judges what is worth saying. When the engine reports that it resolves to a remote endpoint, the Models row says so in words: while voice mode is on, on each message, your last six messages to the agent (scrubbed of code, paths and secrets) and the agent's last spoken line go to that endpoint's host; and for each line the voice is about to say, the part of the agent's reply it comes from (scrubbed, up to 700 characters) and your last request go there too. With voice mode off, or with no classifier connected, nothing is sent. A classifier that runs on this device is shown as such. This pack only reports it: the engine does the sending.
  • Talk live sends your microphone to the live voice's provider for as long as a call is open; the pane says so whenever the engine offers Live. With Codex realtime the browser streams it straight to the provider; the engine holds the call's control channel, which the provider also sends a duplicate of the audio on, and the engine drops that duplicate unparsed and unlogged. With an ElevenLabs agent it goes through the engine, which passes it on unparsed and unlogged.
  • Read-only. The pane reads three public root facts (speech/profiles, agents/list, models) through the root-fenced Store and writes nothing. Choosing an agent's voice, trying a voice and showing which model the voice role resolves to each need a door the Store does not give a root seat yet; the pane says how to do them by hand instead of drawing controls that would not work. The classifier role's resolution is the exception: newer engines publish it as classifier on speech/profiles, and an older engine that omits it leaves the pane silent about it.
  • It shows facts the engine publishes. On an engine that publishes none of them the pane says there is no speech engine, never throws; a malformed row is dropped.
  • The tuning values are not shown. The voice block of config.json is not in the Store, so the pane lists the keys and their defaults, not what you set. Choosing the live voice is likewise by hand, in the voice's .yml.

Build and test

bun run build   # dist/index.mjs
bun run test    # bun test test/

The host loads the committed dist/index.mjs, not src/: rebuild and commit it with every change.