Automatic detection of hyperarticulated speech

Abstract

Hyperarticulation is a speech adaptation that consists of adopting a clearer form of speech in an attempt to improve recognition levels. However, it has the opposite effect when talking to ASR systems, as they are not trained with such kind of speech. We present approaches for automatic detection of hyperarticulation, which can be used to improve the performance of spoken dialog systems. We performed experiments on Let’s Go data, using multiple feature sets and two classification approaches. Many relevant features are speaker dependent. Thus, we used the first turn in each dialog as the reference for the speaker, since it is typically not hyperarticulated. Our best results were above 80 % accuracy, which represents an improvement of at least 11.6 % points over previously obtained results on similar data. We also assessed the classifiers’ performance in scenarios where hyperarticulation is rare, achieving around 98 % accuracy using different confidence thresholds.