Acoustic Feature Analysis and Discriminative Modeling of Filled Pauses for Spontaneous Speech Recognition

摘要

Most automatic speech recognizers (ASRs) concentrate on read speech, which is different from spontaneous speech with disfluencies. ASRs cannot deal with speech with a high rate of disfluencies such as filled pauses, repetitions, lengthening, repairs, false starts and silence pauses. In this paper, we focus on the feature analysis and modeling of the filled pauses "ah," "ung," 'um," "em," and "hem" in spontaneous speech. Karhunen-Loeve transform (KLT) and linear discriminant analysis (LDA) were adopted to select discriminant features for filled pause detection. In order to suitably determine the number of discriminant features, Bartlett hypothesis testing was adopted. Twenty-six features were selected using Bartlett hypothesis testing. Gaussian mixture models (GMMs), trained with a gradient decent algorithm, were used to improve the filled pause detection performance. The experimental results show that the filled pause detection rates using KLT and LDA were 84.4% and 86.8%, respectively. A significant improvement was obtained in the filled pause detection rate using the discriminative GMM with KLT and LDA. In addition, the LDA features outperformed the KLT features in the detection of filled pauses.