Congratulations to @feiyuehchen.bsky.social and Lateef Adeleke who successfully converted their term paper on feature-level ASR evaluation into a publication to appear at the AfricaNLP workshop at EACL next month! pre-print: arxiv.org/abs/2602.04716
Post
arxiv.orgLinguistically Informed Evaluation of Multilingual ASR for African LanguagesWord Error Rate (WER) mischaracterizes ASR models' performance for African languages by combining phonological, tone, and other linguistic errors into a single lexical error. By contrast, Feature Erro...
It features 1) The first ASR baseline for Uneme (endangered Edoid lg) 2) A Tone Error Rate (TER) extension to Feature Error Rate (FER), calculated by carefully scoring phonetic feature vectors 3) Evidence that WER misrepresents ASR performance by collapsing feature-level errors into lexical failure
0 reposts0 likes
The Uneme data was collected and curated as part of Lateef's ongoing fieldwork with the Uneme community in Nigeria. He's working under an ELDP grant and the larger collection is indexed here: www.elararchive.org/uncategorize...
elararchive.orguneme-adeleke-0831 | Endangered Languages Archive