Intent over keywords
Mapping many phrasings onto a few actions is what makes a voice interface feel like it works. The classifier absorbs the variation so the execution layer can stay small and predictable.
All projects
Personal Voice Assistant
Hands-free Spotify control with intent recognition that actually holds up.
Source is private — happy to walk through the architecture on a call.
Hands-free
full playback control
NLP
intent-based, not keyword-matched
A voice assistant for Spotify built around intent recognition rather than rigid command matching.
Why it mattered
This was the project where speech stopped being a gimmick to me. The gap between "transcribes words" and "understands what you meant" is the entire engineering problem, and it is where every voice interface lives or dies.
The problem
Keyword-matched voice control breaks the moment you phrase something naturally. "Play the album" and "put on that record" mean the same thing, and a command table cannot cover the space.
What I built
- Speech-to-text pipeline feeding a Gemini-backed intent classifier rather than a fixed command table.
- NLP intent recognition mapping varied natural phrasing onto a small set of reliable Spotify actions.
- Execution layer against the Spotify API with graceful handling of ambiguous requests.
Engineering notes
Curious about any of this?
I am glad to go deeper on the architecture, the tradeoffs, or the parts that did not work the first time. That conversation is usually more useful than the README.