Build a Wake-Word Detector for a Smart-Speaker Startup
Overview
What this challenge is about.
You receive a small public Japanese-speech dataset, 30 hours of recorded wake-phrase utterances from 50 volunteers, and 200 hours of background-noise recordings. Train a lightweight (about 100k-parameter) keyword-spotting model that fits in 200 KB of RAM. Measure detection rate at the false-accept budget of one per 24 hours of background audio. Benchmark inference latency on a Raspberry Pi 4 proxy (the speaker's CPU class). Deliver the trained model, eval report, and a 2-page engineering memo with rollout caveats.
The Brief
What you'll do, and what you'll demonstrate.
Train an on-device Japanese wake-word detector with strong recall at a strict 1-per-24h false-accept budget.
Earning criteria — what you'll demonstrate
- Train a lightweight keyword-spotting model
- Evaluate detection at a fixed false-accept budget
- Optimize a model for on-device deployment
- Document rollout risks for a hardware-launching team
Program Fit
Where this fits in your program.
Sharpens the same skills your degree expects you to demonstrate.
Skills
Skills you'll demonstrate.
Each one shows up on your verified credential.
Careers
Roles this prepares you for.
Real titles. Real skill bridges. Pick the one closest to your trajectory.
NLP Engineer
On-device keyword spotting with strict false-accept budgets is the daily reality of NLP/speech engineers at smart-speaker and wearables companies.
This challenge sharpens
- keyword-spotting
- speech-recognition
- on-device-ml
Machine Learning Engineer
Tight model-size and latency constraints translate directly to MLE work on edge-deployed models broadly.
This challenge sharpens
- on-device-ml
- model-compression
- evaluation
AI Engineer
Wiring a research-grade model into a hardware-rollout memo is AI-engineer glue work at consumer-AI startups.
This challenge sharpens
- pytorch
- on-device-ml
- model-compression