Computer analysis of children's non-native English speech for language learning and assessment

Qian, Mengjie ORCID: 0000-0003-2614-5411 (2021). Computer analysis of children's non-native English speech for language learning and assessment. University of Birmingham. Ph.D.

Preview

Qian2021PhD_Redacted.pdf
Text - Redacted Version
Available under License All rights reserved.
Download (10MB) | Preview

Abstract

Children's ASR appears to be more challenging than adults' and it's even more diﬃcult when it comes to non-native children's speech. This research investigates diﬀerent techniques to compensate for the eﬀects of non-native and children on the performance of ASR systems. The study mainly utilises hybrid DNN-HMM systems with conventional DNNs, LSTMs and more advanced TDNN models. This work uses the CALL-ST corpus and TLT-school corpus to study children's non-native English speech.

Initially, data augmentation was explored on the CALL-ST corpus to address the lack of data problem using the AMI corpus and PF-STAR German corpus. Feature selection, acoustic model adaptation and selection were also investigated on CALL-ST. More aspects of the ASR system, including pronunciation modelling, acoustic modelling, language modelling and system fusion, were explored on the TLT-school corpus as this corpus has a bigger amount of data. Then, the relationships between the CALL-ST and TLT-school corpora were studied and utilised to improve ASR performance.

The other part of the present work is text processing for non-native children's English speech. We focused on providing accept/reject feedback to learners based on the text generated by the ASR system from learners' spoken responses. A rule-based and a machine learning-based system were proposed for making the judgement, several aspects of the systems were evaluated. The inﬂuence of the ASR system on the text processing system was explored.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):