Comprehensive literature review on children automatic speech recognition system, acoustic linguistic mismatch approaches and challenges

Automatic Speech Recognition (ASR) system for children is as important as for adults since children are more dependent on these systems nowadays, such as computer games, reading tutors, foreign language learning tools, etc. Consequently, this article aims to present several important aspects related to children's speech recognition systems, in which a comprehensive review is presented. Acoustic and linguistic challenges of children's speech are presented thoroughly to understand the basic anatomy of children's articulation organs. A variety of challenges exist for the development of children's ASR, such as the collection of children's speech data is a very complex task; the available child corpora are not publicly accessible, children's speakers differ greatly due to linguistic and acoustic variations, and ASRs developed for one age group are not suitable for another age group. All these challenges are systematically described in this article. Various data augmentation methods are also explored here, along with different approaches to develop ASR in children's speech. It has been observed that the inaccessibility of child corpora publicly is a significant barrier to children's ASR. Apart from the challenges mentioned earlier related to children’s ASR, an attempt has been made to thoroughly review the children’s ASR in the case of Punjabi language, as this language is ranked 10th most spoken globally and is still considered a low-resource language. Further, various approaches for the development of children’s ASR such as traditional, hybrid and end-to-end (E2E) networks are also reported. In addition, an analytical summary and discussion are included.