@memo-air/voice

[!WARNING] Experimental: the voice layer’s API may change in any 0.x minor release.

The huddle’s voice layer on LiveKit Agents for Node: a hidden ear per person, one speaking floor, barge-in, an echo filter, and a talker per agent that hands long work to the agent’s harness. It takes the agents, a router, a worker and keys as arguments and knows nothing of the app around it.

Main exports

• createVoice(setup, host) — returns join, leave, isAvailable and close. setup gives the LiveKit URL, a token minter, the keys and the talker model; host gives agentsIn, route, work, postToChat, writeLine and optionally postNotice and voiceCatalog.
• voiceFor, defaultVoices — a stable, distinct voice per agent.
• speechProviderOf — ElevenLabs, else OpenAI, else none.

Used by

apps/airlink/server/main.ts. See docs/huddle.md.

Licence note

It depends on @livekit/local-inference@0.2.7 for on-device turn detection: Apache-2.0, with model weights under the LiveKit Model License. It is a pinned exception in scripts/licenses.mjs.

[!NOTE] Not yet published to npm. It installs from this repository’s workspaces; publishing is coming.