Prescriptive Persistence
Quantifying the breakdown in human-AI pedagogical co-regulation in feedback on English language learners’ writing. Introduces ΔP, a measure of how persistently AI feedback pushes learners toward prescriptive norms.
Dialect-Aware Speech Recognition
Post-ASR divergence evaluation and model-agnostic adaptation for dialectal, multilingual, accented, code-switched, and disfluent speech.
In progress
Code-Switching & Multilingual ASR
How code-switching and multilingual language use affect automatic speech recognition performance and how that performance gets evaluated.
In progress
Why evaluation matters
Standard accuracy scores count mismatched words. They can’t tell a system error from a dialect the system was never built to understand.
One spoken sentence
Scored as 2 errors in 4 words
| What was saidreference | he | be | working | late |
|---|---|---|---|---|
| What the system wrotehypothesis | he's | — | working | late |
| Scored asWER operation | changed | dropped | ✓ | ✓ |
In African American English, habitual be means he usually works late. The system wrote he’s working late, meaning right now. Word error rate (WER) counts two mistakes and stops there. It has no way to register that the meaning changed. My research builds evaluation that can.