A method for unsupervised detection of the boundary of coarticulated units from isolated speech using recurrence plot

Abstract

One of the major step for Automatic Speech Recognition (ASR) is to mark the boundary of two consecutive co-articulated units. While attempt has been done to use cross recurrence plot (supervised learning) to address similar problems [1], here we propose an unsupervised approach using Recurrence Plots (RP) to address the same problem. The novelty of the work is two fold. First, we report a novel approach on using RP to identify co-articulated boundaries through prominent visual patterns. Second, we use a different quantitative approach rather than usual Recurrence Quantification Analysis (RQA) matrix [2] for automatic detection of transition boundaries. The proposed algorithm is applied on isolated spoken numerals in Bangla, a major Indian language. The results obtained from a considerably large database shows that the proposed method is a potential candidate to address the problem of Co-articulated Units Boundary (CUB) detection. © 2011 Springer-Verlag.