THE HACKATHON DESK · ★ FIRST PLACE
PREDICTIVE SPEECH CASCADE TAKES FIRST PLACE
An AAC communication system that learns how you speak — and drafts the sentence before you finish tapping
SEATTLE — Non-verbal users rely on augmentative and alternative communication devices to speak, but most commercial tools are static: they do not learn individual patterns, context, or preferences over time. Users must build every phrase from scratch, and generic suggestions rarely match what they actually mean to say. ConnectAble, which took first place this season, was built to close that gap.
The system pairs a 66-symbol communication board — five categories spanning food, feelings, actions, places and people — with a personalized prediction engine that retrains itself nightly from the user's own phrase history. As each symbol is tapped, a top bar offers whole utterances, not letters: a bet that context beats keystrokes.
HOW IT WORKS
Predictions run through a three-layer cascade ordered by cost. Layer one is an instant bigram lookup against a local vocabulary store. Layer two performs semantic vector search over past phrases via ChromaDB. Layer three falls back to a local phi3 model served by Ollama when the phrase database is sparse.
Every night at 2 a.m. the system retrains from logged history, rebuilding the bigram map and re-embedding phrases so the model improves without user intervention. Speech is produced offline through pyttsx3, with ElevenLabs streaming available for users who want a more natural voice. All phrase data rests in an AES-256 encrypted SQLite database, and the language model runs locally — nothing leaves the device by default.
LESSONS FROM THE FLOOR
The ordering of the cascade proved decisive: putting the cheapest operation first kept suggestions instant for common phrases while preserving quality for novel ones. And for assistive technology, the builder notes, communication history is among the most sensitive data a person produces — privacy was treated as a core feature, not an afterthought.
Source available → github.com/SNBest1/ConnectAble ↗
CONTINUED: Offline AI watches over Alzheimer's care, Page A2 →