Native phone-class geometry improves second-language pronunciation analysis
A Native-Reference Phone-Class Geometry for Second-Language Pronunciation Analysis
Computation and LanguageSound
Summary
Assessing how well someone speaks a second language is hard because automatic systems often give only a general score without explaining pronunciation errors. The authors created a way to compare a learner’s speech to native speakers using a special 'geometry' built from recordings of native speech. This method can measure pronunciation quality without needing exact text matches or labels, and it works on spontaneous speech. Their tests found this approach reflects speaking skill and pronunciation quality fairly well.
What this means in practice
- •For language learning app developers: Integrate a pronunciation scoring method that analyzes learner speech without needing exact native text matches or labels, enhancing feedback quality.$Commercial implications: Enables creation of scalable pronunciation assessment features for language learning apps that work on spontaneous speech.
- •For speech therapy software teams: Use native-reference phone-class geometry to measure pronunciation deviations in second-language learners, supporting speech improvement tracking without scripted prompts.
Authors
Tina Raissi, Nhan Phan, Chenxiao Wang, Mikko Kurimo
Abstract
Automatic speaking assessment systems can provide holistic proficiency scores, but often lack interpretable measures that characterize pronunciation quality. We propose a native-reference phone-class geometry for measuring second language (L2) pronunciation deviation without requiring pronunciation labels, read-aloud prompts, or matched recordings of the same text from native and L2 speakers. Given a native speech corpus, we average frame-level self-supervised representations for each context-dependent phone-class and use singular value decomposition (SVD) to derive a compact native-reference coordinate system. For each L2 utterance, we compute the corresponding averages and project them into the native-reference space. We then demonstrate that the distances between L2 and native-reference coordinates for matched phone-classes show consistent negative correlations with holistic speaking proficiency on the Dev subset of the Speak and Improve Corpus 2025 (Spearman's $ρ\!=\!-0.53$) and with pronunciation quality on the learner subset of the English Read by Japanese Students dataset ($ρ\!=\!-0.34$). These findings suggest that the proposed geometry captures acoustic-phonetic information relevant for proficiency rating while remaining applicable to spontaneous L2 speech without matched native recordings.