Enhancing Automatically Discovered Multi-level Acoustic Patterns Considering Context Consistency With Applications in Spoken Term Detection

摘要

This paper presents a novel approach for enhancing the multiple sets ofacoustic patterns automatically discovered from a given corpus. In a previouswork it was proposed that different HMM configurations (number of states permodel, number of distinct models) for the acoustic patterns form atwo-dimensional space. Multiple sets of acoustic patterns automaticallydiscovered with the HMM configurations properly located on different pointsover this two-dimensional space were shown to be complementary to one another,jointly capturing the characteristics of the given corpus. By representing thegiven corpus as sequences of acoustic patterns on different HMM sets, thepattern indices in these sequences can be relabeled considering the contextconsistency across the different sequences. Good improvements were observed inpreliminary experiments of pattern spoken term detection (STD) performed onboth TIMIT and Mandarin Broadcast News with such enhanced patterns.