STARK

STARK

Async, low-latency Python framework for speech interfaces.Pre-GPT context-aware multilingual NLU system.Extended with ML pipelines, LLM integration, and a custom agentic engine for autonomous workflows.
voicefor developersopen source

STARK (Speech And Text Algorithmic Recognition Kit) is our voice interface development platform: offline, embeddable, and built to give products their own voice without shipping your users' audio to someone else's cloud.

We built it because MajorDom needed a voice layer that respected privacy and worked without internet. It grew into a platform developers can build on, and one we license to companies that want custom voice interfaces in their own products. First commit: 2020. It now also powers Archie.

What it can do

  • Mid-sentence language switching. Parse commands that flip between languages in the same utterance, no extra setup.
  • Multiple commands, one sentence. Pulls out and runs several distinct commands from a single thing someone said.
  • Background tasks with live progress. Kick off long operations that respond immediately, keep listening, and report progress, cancel included.
  • Fuzzy, phonetic matching. Handles accents, misspellings, and cross-language names deterministically, not by guessing with a model.
  • Stateful, nested conversations. Multi-level menus and follow-ups that actually remember where you were.
  • No AI required, but never blocked either. Core recognition runs on the powefull pattern matching; plug in an LLM as a fallback if you want one, not because you have to.
  • Swap the whole interface. Voice, text, a Telegram bot, API, or something custom, without touching your command logic.
  • Fully offline, minimal footprint. Four required dependencies, everything else optional, no API keys, all data stays on-device.
  • AI Agent Platform (coming in v5). Agents built on the same async/background command infrastructure, fully backward compatible.

Docs & source

Licensing

STARK is free for non-professional non-commercial home use by enthusiasts, teachers, students, and the DIY community. Any other use requires a license, including internal business tools, expo demos, and prototypes that don't directly generate profit. Terms are discussed individually, and under some conditions a license can be granted without payment, for example for short-term demos or within a partnership or a contract where we build on STARK for you.

We also take on contract work building directly on STARK, including adding support for any language you need, even ones we don't ship out of the box.

Work with us: Commercial licensing

License STARK for custom voice interfaces in your products.

Work with us: Contract development on STARK

We build directly on STARK under contract, including adding support for any language you need, even ones we don’t ship out of the box.