Detection of Vowel Offset Points Using Non-Local Similarity Between Speech Samples

摘要

Automatic detection of vowels is not only an important but also a challenging problem. Vowel offset point (VEP) is the instant of ending of a vowel. Like vowel onset points (VOPs), VEPs are equally important for accurate marking of vowels and analysis of speech signal. The transition in the signal magnitude at the VEPs is quite different when compared to the VOPs. Consequently, most of the front-end features proposed for the detection of VOPs fail to detect the VEPs. Performance of the existing features also reduces significantly in the case of noisy speech signals. In this work, a robust frontend speech parametrization approach is proposed for enhancing the discrimination at the VEPs. In the proposed approach, weight values are assigned to each of the sample points by computing the similarity present in the samples belonging to two different frames within a search neighborhood. The weight values (WVs) computed from the non-local similarity (NLS) is significantly less when the frames under consideration are similar in comparison to the dissimilar ones. Since the vowels are longer regions and exhibit periodicity, there will be more similarity in the case of frames belonging to the these regions. On the other hand, the frames belonging to the non-vowel regions and noises will be dissimilar. In this work, WVs computed from the NLS is used as a feature for detecting the VEPs in a given speech signal. The proposed method is observed to outperform the deep neural network - hidden Markov model based classifier under both clean and noisy test conditions even after the inclusion of a recently proposed speech enhancement module.