Accurate phoneme segmentation method using combination of HMM and Fuzzy Inference system

The aim of this study, is to improve the accuracy of automatic segmentation. In the last twenty years, manual speech segmentation is always considered as the most accurate method for speech segmentation. However, this is a handwork by experts and a time-consuming work. Compared with this, automatic segmentation methods are much stable and faster. Unfortunately, achieving accurate segmentation is still a challenging task. Recently, some researches attempt to improve the accuracy of automatic segmentation by using some statistical correction procedures or learning methods on HMM-based forced-alignment. The refinement for HMM-based forced alignment in automatic speech segmentation is still not accurate enough. This paper presents an effective approach based on an adaptive neuro fuzzy inference system (ANFIS) for refining the output of the traditional HMM. Compared with other competitive methods, such as SVM, handmade fuzzy-logic and linear method, ANFIS has advantages in dealing with the problem of non-linear, fuzzy and can be trained in a completely automatic way. This study combined ANFIS instead of linear system used in the state-of-the-art research with HMM-based forced alignment to improve the automatic phoneme segmentation. The results show that ANFIS successful learn the rules of manual speech segmentation strategies and can significantly improve forced alignment accuracy. The proposed system already achieved 91.6% agreement within 20 msec for manual segmentation on the TIMIT corpus, comparing the 89.98% for the linear system used in the outstanding work.