← All posts
STANDARD.SITE

Optional Speech: Gemini Live Voice for the Web

An opt-in speech layer using Gemini Live. The model gains a keating_voice tool for short utterances while reasoning stays in the normal text loop.

featureversion:0.3.1

Keating now has an optional speech layer. The teacher still thinks, verifies, plans, and steers through the normal model/tool loop, but it can hand short learner-facing moments to a dedicated voice tool when speech is useful.

How It Works

  • Opt-in by design: Speech stays disabled until you turn it on. In the web app, use the speaker button in the chat header.
  • Voice tool: When enabled, the model gets keating_voice for short questions, redirects, recaps, and encouragement.
  • Gemini Live path: The browser route uses gemini-3.1-flash-live-preview with the Google API key stored in Settings.
  • Normal model remains in charge: Reasoning, verification, and correction still happen through the regular teaching tools instead of being hidden inside the voice layer.

Shell Support

The CLI config now includes a disabled-by-default speech block. When enabled, the Pi extension registers keating_voice and emits transcript-safe voice tags:

[voice voice=conversational tags=question,verify pace=normal affect=curious] What would you expect to happen next?

The first shell version is provider-neutral and tag-based. The web version is the first audio-backed path.