Robust Multi-Keyword Spotting of Telephone Speech Using Stochastic Matching

Abstract

In telephone speech recognition, the acoustic mismatch between the training and the test environment often causes severe degradation due to the channel distortion and ambient noise. In this paper, a two-level codebook-based stochastic matching (CBSM) is proposed to deal with the acoustic mismatch. For multi-keyword detection, we define a keyword relation table and a weighting function for reasonable keyword combinations. In the multi-keyword spotting system, 94 right context-dependent INITIAL's, 37 context-independent FINAL'S and 1 silence model are adopted. In order to evaluate the multi-keyword spotting system, 1275 faculty names and department names are selected as the keywords. Using a testing set of 2400 conversional speech utterances from 8 speakers, the proposed two-level CBSM can reduce the recognition error rate from 36.52% to 13.4%.