Speech Recognition with Limited Resources for Children and Adult Speakers

Children are natural users of speech technology. However, children speech has been proved to be harder than adults. This implies that in order to build a children speech recogniser that is comparable with an adult recogniser it would be necessary a larger amount of recordings. In this paper, we present our experiments building such recognisers but with a limited amount of resources, this is a corpus with few speakers and recordings. We explore the effects of combining both resources in different ways. We also show the performance of the recognisers in a real life scenario with spontaneous speech.