Overview
Lisper is a lisp-focused AI speech-practice app built for the Kaggle Gemma 4 Good Hackathon. The product loop is intentionally narrow: record a phrase, get concise feedback about likely /s/ and /z/ articulation patterns, then try again.
How It Works
- The app centers lisp practice instead of treating lisps as a small part of generic speech therapy.
- The release includes a LoRA adapter, merged model, browser-ready ONNX/WebGPU package, public demo, and writeup.
- The browser path is designed for local, keyless inference rather than requiring a hosted closed-model API.
Hard Parts
The important product constraint is tone. The app should give useful feedback without pretending to replace a speech-language pathologist, and it should make repetition feel low-pressure rather than clinical.
Results
The hackathon release shipped public code, model artifacts, and a demo surface. The final v18 hybrid eval passed on 2,000 held-out rows with 0 hard errors, and the browser target is the q4f16 ONNX/WebGPU package.
