The Korean Speech Recognition Sentences: A Large Corpus for Evaluating Semantic Context and Language Experience in Speech Perception

Abstract

Purpose: The aim of this study was to develop and validate a large Korean sen-tence set with varying degrees of semantic predictability that can be used for testing speech recognition and lexical processing. Method: Sentences differing in the degree of final-word predictability (predict-able, neutral, and anomalous) were created with words selected to be suitable for both native and nonnative speakers of Korean. Semantic predictability was evaluated through a series of cloze tests in which native (n = 56) and nonnative (n = 19) speakers of Korean participated. This study also used a computer lan-guage model to evaluate final-word predictabilities; this is a novel approach that the current study adopted to reduce human effort in validating a large number of sentences, which produced results comparable to those of the cloze tests. In a speech recognition task, the sentences were presented to native (n = 23) and non-native (n = 21) speakers of Korean in speech-shaped noise at two levels of noise. Results: The results of the speech-in-noise experiment demonstrated that the intelligibility of the sentences was similar to that of related English corpora. That is, intelligibility was significantly different depending on the semantic condition, and the sentences had the right degree of difficulty for assessing intelligibility differences depending on noise levels and language experience. Conclusions: This corpus (1,021 sentences in total) adds to the target lan-guages available in speech research and will allow researchers to investigate a range of issues in speech perception in Korean.