Research program
Four connected research directions
My work connects documentary fieldwork, linguistic analysis, corpus development,
and computational modeling. These areas are not separate projects; they form a
common research program focused on how underdescribed languages can inform both
linguistic theory and language technology.
01 · Speech Technology
Low-Resource Speech Technology
How can speech systems learn robustly from small, heterogeneous,
and linguistically complex datasets?
I build and evaluate automatic speech recognition systems for languages
with limited transcribed speech resources. My current work investigates
cross-lingual transfer, data-efficient adaptation, speech-style mismatch,
naturalistic versus constrained speech, tonal contrasts, and the limits
of multilingual pretraining.
A central concern is ecological validity: models that perform
well on carefully elicited or read speech may behave very differently on
spontaneous documentary recordings. I therefore treat recording style,
speaker variation, community variation, and linguistic structure as central
experimental variables rather than noise to be removed.
ASR
Whisper
XLS-R
Cross-lingual transfer
Naturalistic speech
Style robustness
02 · Model Evaluation
Linguistically Informed AI Evaluation
What do aggregate metrics hide about what speech and language models
actually learn?
Word Error Rate and Character Error Rate are useful, but they can collapse
linguistically different error types into a single score. Our work develops
and applies evaluation approaches that ask which phonological, tonal,
morphological, and grammatical distinctions are preserved or lost by a model.
This direction is represented by our recent AfricaNLP 2026 work on multilingual ASR
for African languages, where phonological-feature and tone-sensitive evaluation
reveals systematic model behavior that word-level accuracy alone cannot show.
Read the AfricaNLP 2026 paper →
WER / CER
Phonological features
Tone
Error analysis
African NLP
03 · Grammar & Variation
Corpus-Accountable Grammar & Variation
How can grammatical description capture the variation that actually
occurs in documentary corpora?
I study tense, aspect, negation, grammatical tone, focus, movement, and
clause structure in underdescribed African languages, especially Edoid
languages.
I am particularly interested in the relationship between targeted elicitation
and naturalistic evidence, and in grammatical descriptions that remain
accountable to patterns across speakers, communities, varieties, and
discourse contexts.
My work treats variation not as a defect in the dataset, but as an empirical
fact with consequences for linguistic theory, documentation practice, and
computational modeling.
Tense & aspect
Negation
Grammatical tone
Focus
Clause structure
Variation
04 · Research Infrastructure
Documentation-Informed Language Technology
How can documentary resources become computationally useful without
stripping away linguistic richness?
For many underdescribed languages, computational research cannot begin with
a large downloaded benchmark. It begins with field relationships, recording
decisions, transcription conventions, annotation, lexical analysis, and
language-specific knowledge.
I therefore work across the full pipeline: audiovisual documentation,
transcription and annotation, corpus and lexicon development, grammatical
analysis, speech-dataset creation, model training, and linguistic evaluation.
This creates a feedback loop in which documentation informs technology,
while model errors generate new questions about language structure,
representation, and variation.
Fieldwork
Corpus development
ELAN
FLEx
Speech datasets
Model evaluation