Child Speech Recognition on End-to-End Neural ASR Models

Automatic speech recognition (ASR) is an area which is having maximum improvement and is developing to a very high extent as far as adult speech is concerned. Nevertheless, child speech recognition is an area which is least explored. This is because of the scarcity of a neatly labelled corpora. Precisely, to train a speech recognition model from scratch and to make it work child speech is a challenging task as the availability of child speech data is limited. In this context, exploring child speech recognition can also help us improve the current speech recognition systems since on validating it in such a manner that it will be useful for child speech may help the systems work in a good way such that we will be able to validate it for everyone irrespective of the age variation in which a speech utterance is produced. State-of-art models in this domain perform various complex speech recognition tasks such as machine translation and ASR on multilingual datasets as well as it is useful in performing self supervised speech recognition tasks. This process has proved to work well for speech data which consists of recordings with adult data and it is not yet validated on child speech. So the present work tries to validate the performance six state-of-the-art ASR models on children speech and report the corresponding word error rate.