The effects of child language development on the performance of automatic speech recognition

[摘要] In comparison to adults’, children’s ASR appears to be more challenging and yields inferior results. It has been suggested that for this issue to be addressed, linguistic understanding of children’s speech development needs to be employed to either provide a solution or an explanation. The present work aims to explore the influence of phonological effects associated with language acquisition (PEALA) in children’s ASR and investigate whether they can be detected in systematic patterns of ASR phone confusion errors or they can be evidenced in systematic patterns of acoustic feature structure. Findings from speech development research are used as the framework upon which a set of predictable error patterns is defined and guides the analysis of the experimental results reported. Several ASR experiments are conducted involving both children’s and adults’ speech. ASR phone confusion matrices are extracted and analysed according to a statistical significance test, proposed for the purposes of this work. A mathematical model is introduced to interpret the emerging results. Additionally, bottleneck features and i-vectors representing the acoustic features in one of the systems developed, are extracted and visualised using linear discriminant analysis (LDA). A qualitative analysis is conducted with reference to patterns that can be predicted through PEALA.

[发布机构] University:University of Birmingham;Department:School of Engineering, Department of Electronic, Electrical and Systems Engineering

[关键词] P Language and Literature;P Philology. Linguistics