Uneme Documentation Project
ELDP-funded · Principal Investigator
A multi-year documentation project creating naturalistic audiovisual recordings, elicitation, annotations, grammatical analyses, lexical resources, and archival deposits focused on Uneme language, cultural practices, oral history, and indigenous knowledge.
Speech Corpus & ASR
The research examines the effect of speech-styles, training composition, cross-lingual transfer, and data-efficient adaptation in low-resource ASR.
Uneme–English Lexical Resources
A multimedia lexical resource integrating field recordings, elicited lexical data, grammatical information, and corpus examples. The workflow uses FLEx alongside documentary annotations and is designed to support both community-oriented dictionary outputs and linguistic research.
Linguistically Informed ASR Evaluation
Methods for analyzing speech-recognition behavior using language-specific phonological structure, tone, and contrastive features rather than relying only on aggregate word-level scores.
Grammar & Variation in Edoid Languages
Fieldwork- and corpus-based research on tense, aspect, negation, grammatical tone, focus, movement, and morphosyntactic variation in Uneme and related Edoid languages.
Methods & Tools
Computational
Speech & Evaluation
Linguistic Data
Selected Research Outputs
Public datasets, code, archival collections, and publications from my current research program.
Documentation & Archive
Uneme Documentation Project
ELDP-funded documentation of Uneme language, indigenous iron technology, oral history, cultural practices, and linguistic variation across Uneme-speaking communities.
Code & Experiments
Uneme ASR Style Generalization
Reproducible experiments investigating naturalistic versus constrained speech, training composition, cross-lingual transfer, and style robustness in low-resource Uneme automatic speech recognition.
Publication
Linguistically Informed Evaluation of Multilingual ASR for African Languages
AfricaNLP 2026 work on Yorùbá and Uneme showing how phonological-feature and tone-aware evaluation can reveal model behavior hidden by aggregate word-level metrics.