Recognition of read and spontaneous children's speech using two new corpora

Abstract

This paper describes some of the results of research into automatic recognition of children’s speech which has been conducted as part of the European Framework 5 ‘PF STAR’ project. Two new corpora of British English children’s speech are described. The first comprises over 14 hours of read data from 159 children, while the second includes 1 hour and 23 minutes of spontaneous and emotional speech from 30 children. A partition of the data into training, evaluation and test sets is proposed, and the results of ‘baseline’ speech recognition experiments are presented. The results fail to demonstrate a significant improvement from the use of age dependent acoustic models, or that the emotional speech is more difficult to recognise than ‘ordinary’ spontaneous speech.