Parsing "play Radiohead" without a neural net

What it does

Adapt turns spoken or typed natural language into structured intents using hand-defined entities and rules. You register vocabulary (“play” is a Listen Command, “Radiohead” is an Artist) and build intents from required or optional pieces. It’s the parser behind the Mycroft open-source voice assistant.

The interesting bit

The framework is deliberately non-neural. You define what matters — no training data, no GPU, no stochastic surprises. The Pandora example in the README shows dynamic vocabulary populated from an API at runtime, which is where this gets useful: static commands plus live data.

Key highlights

• Rule-based matching with require() and optionally() constraints
• Runtime vocabulary registration (good for user-specific data like playlists or station names)
• Debug serialization via adapt.tools.debug for reproducible issue reports
• Apache 2.0 licensed, though the project requires a CLA for contributions
• Production-ready badge, but Travis CI and Coveralls links suggest older tooling

Caveats

• README examples use Python 2-era virtualenv syntax and Travis CI (deprecated), so setup may need adjustment
• No performance benchmarks or latency claims in the documentation
• Requires manual vocabulary curation; scaling to broad domains means more elbow grease

Verdict

Good fit if you’re building a voice interface with a constrained, known vocabulary and want debuggability over magic. Skip it if you need open-domain understanding or expect users to say things you haven’t explicitly planned for.