Hypo and Hyperarticulated Speech Data Augmentation for Spontaneous Speech Recognition

摘要

Among many challenges in spontaneous speech recognition, we focus on the variability of speech depending on the degree of articulation such as hypo and hyperarticulation. In this paper, we investigate the feasibility of the past acoustic-phonetic studies on the variability of speech in terms of the data augmentation of a spontaneous speech recognition system. To do so, we develop data augmentation approaches to reflect the acoustic-phonetic characteristics of hypo and hyper-articulated speech. Since our approaches are based on signal processing methods they do not require a model learned from supervised or unsupervised data. A series of speech recognition tests are conducted across various speech styles. The results show that we are able to achieve meaningful performance gain by using our approaches. It also indicates that the past acoustic-phonetic knowledge of the variability of speech is useful for improving the recognition performance of spontaneous speech including hypo and hyper-articulated speech.