Research initiative
Inside the folder
Researcher, Internal and Part Time
Microsoft · September 2025 - Present
Speech evaluationFairnessCode-switchingCalibration
Description
Evaluation work that separates natural language-transfer patterns from speech pathology in code-switched, noisy African speech.
Ongoing internal, part-time research
The research question
How can a speech-evaluation system avoid mistaking language transfer or code-switching for impairment?
My contribution
- Built an evaluation pipeline to distinguish linguistic variation from speech impairment
- Measured word error rate, phoneme F1, calibration, and false-positive disparities
- Contributed to dataset design, analysis, and manuscript development within an interdisciplinary team
Outcomes & evidence
- An evaluation pipeline focused on distinguishing language variation from impairment
- Analysis spanning recognition errors, phoneme-level performance, calibration, and disparity in false positives
- Ongoing dataset and manuscript contributions within a five-person interdisciplinary team
Approach & methods
- Evaluated noisy, code-switched African speech rather than assuming a single standardized language variety.
- Used WER and phoneme F1 to examine recognition and phoneme-level errors, alongside calibration and false-positive disparities.
- Connected dataset design with analysis so language variation and potential impairment could be considered separately.
Scope & limitations
- Internal findings and unpublished subgroup measurements are not included here.
- This page describes research evaluation and does not make clinical diagnostic or treatment claims.