Development of a topic-related sentence corpus for speech perception research

Abstract

A large sentence corpus has been developed for use in speech recognition research. Sentences (n=881, three scoring words per sentence) were developed under 23 topics. In the first phase of development subjects rated each individual scoring word for relatedness to its given topic on a Likert scale. Next, two groups of young, normal-hearing listeners (n=16/group) listened and responded to the recordings of the sentences (spoken by a female talker) presented with one of two types of maskers: steady-state noise (S:N=−13 dB) or two other females speaking random sentences (S:N=−8 dB). Each subject responded to half of the sentences with topic supplied and half with no topic supplied. Data analyses focused on addressing two questions: whether supplementation of topic would be more important in the presence of the speech masker versus the noise masker, and how the degree of relatedness of each key word to the topic influenced the effect of topic on recognition. The data showed little difference in how beneficial the topic was for speech versus noise maskers. Moreover, there was a complex relationship between effect of topic, type of masker, and position of the word in the sentence. [Work supported by NIDCD DC01625.]