Assessing
child growth in terms of speech and language development is a critical
indicator of long term learning ability and life-long development progress.
The earlier a child who is at-risk is identified, the earlier support can be
provided to reduce the social impact of the speech or language issue. The
preschool classroom provides an opportunity for monitoring growth in young
children’s interactions. To date, limited research has been possible for
young child based speech recognition in classroom settings due to speech data
access, as well as limitations on speech recognition performance for
naturalistic child communication. This study addresses American English
speech recognition for children’s speech in a naturalistic noisy early
childhood setting, where child age varies from 3 to 5 years. This study
investigates the effectiveness of data augmentation techniques to improve
both language and acoustic models, since this is relatively under explored
for young child speech. We consider alternate text augmentation approaches
using adult data, Web data, and text generated by recurrent neural networks.
We also compare several acoustic augmentation techniques including: speed
perturbation, tempo perturbation, and adult data. In the study, we also comment
on child word count rates to assess child speech development. Finally,
insights are provided into the statistical patterns of naturalistic child
speech such as word complexity, stop words, part of speech, etc., which are
intended to serve as a representative of high quality language engagement in
adult–child learning environments.
