CS 250
Accent Coach
Clinical pronunciation scoring for a second-language nurse: voice drills in, phoneme-level scores out, coaching feedback in between. Consumer dictation tools erase accents by design; this pipeline measures where speech actually differs.
- For
- A working RN whose first language is Korean
- Format
- Telegram voice pipeline, Azure Speech scoring, AI coaching layer
The problem
A registered nurse needed accent reduction for clinical communication. Two tools failed her by design: consumer dictation auto-corrects toward native speech, so it hides exactly the differences a learner needs to see, and generic pronunciation apps have no idea what a nurse actually needs to say.
What I built, by directing AI
A voice-to-coach pipeline: voice notes sent over Telegram, converted and scored through Azure's Pronunciation Assessment at phoneme and IPA level, then an agent coaching layer that turns raw scores into targeted drill feedback.
Around the engine sits a Korean-L1 medical drill bank: ten sections, roughly forty drills, built around the vocabulary used on a hospital shift, not tourist phrases.
Guardrails
Audio never passes through iOS or consumer dictation. The scoring key and all configuration live locally, never in shared notes or chats. The coaching agent runs as a separate profile with its own trust boundary, and the learner controls what is saved.
How I measured it
Every drill session logs an overall score plus per-phoneme results. The baseline is captured: an early assessment scored 77.8 overall, with fluency already at 97 but specific drug-name phonemes scoring as low as 8 and 57. Repeated drills make the improvement curve measurable over time.
What happened
Working end to end and verified live: test clips score correctly, problem words are isolated at the phoneme level, and the learner is using it. Data is collecting; outcome claims wait for the data.
What I'd do next
Weekly trend reporting, a drill bank that grows from her real dictation errors, and connecting practice streaks to frequency.