Boundary detection using continuous wavelet analysis

TL;DR: A novel methodology of automatic and unsupervised boundary detection and classification based on the continuous wavelet transform (CWT) technique is presented.

Abstract: Unsupervised boundary detection and classification is both a theoretically interesting question and an important challenge for speech technology. Theoretical interest lies in exploring how and to what extent is the boundary information encoded in purely acoustic material. For technology, automatic boundary detection facilitates cheap and fast labeling of large corpora of speech data. In this work we present a novel methodology of automatic and unsupervised boundary detection and classification based on the continuous wavelet transform (CWT) technique. Several approaches using lines of minimal amplitude, phase information and wavelet-based estimation of speech tempo are evaluated and compared on Boston Radio News Corpus data. The results show that this methodology using hierarchical information encoded in speech signal compares favorably with traditionally used supervised boundary detection techniques using acoustic information.