APPLICATION OF VTS TO ENVIRONMENT COMPENSATION WITH NOISE STATISTICS.pdf

APPLICATION OF VTS TO ENVIRONMENT COMPENSATIONWITH NOISE STATISTICSNam Soo Kim Do Yeong Kim Byung Goo Kong Sang Ryong KimHuman Computer Interaction Lab.Information Processing SectorSamsung Advanced Institute of TechnologyP. O. Box 111, Suwon 440-600, Koreanskim@green.sait.samsung.co.krABSTRACTRecently, the vector Taylor series (VTS) approach wasproposed as an ecient method for robust speech recog-nition under various environmental conditions. The VTSapproach makes an approximation to the speech contami-nation procedure by a linearized model and estimates theparameter values using the expectation-maximization (EM)algorithm. In this paper, we apply the VTS approach to en-vironment compensation with assumed noise statistics. Inaddition, we present a Bayesian adaptation technique withwhich we can incorporate the a priori knowledge about thenoise statistics to the parameter estimation procedure.1. INTRODUCTIONEnvironment compensation techniques which compensatethe eects of added noise and channel distortion have beendeveloped for robust speech recognition [1] - [3]. In gen-eral, these techniques make assumptions on both the cleanspeech distribution and the environmental contaminationprocedure, and by estimating the environmental parameter-s, transform the noisy speech features to the clean features.Among them, the vector Taylor series (VTS) approach isconsidered to be eective since it can approximate the high-ly nonlinear contamination procedure by a simplied linearmodel, which improves the mathematical tractability in pa-rameter estimation [1], [2].In the VTS approach, the clean speech features are char-acterized by a mixture of Gaussian distributions, and thecontamination procedure is separately approximated foreach mixture component by using the truncated VTS ex-pansion. Environmental parameters such as the added noisefeature and spectral tilt vectors are derived through theexpectation-maximization (EM) algorithm which iterativelyupdates the parameter estimates.I