Essv 2012 C Ontinuous Speech Recognition Using C Orrelation Features and Structured Svm Probability Output

One potential area for improvement in continuous speech recognition is the modelling of phoneme transitions (not transition probabilties) arising from the non-stationarity of speech: refined models can then be used to compute probability distributions which can serve as emission probabilities for HMM-based speech recognition systems. In this paper we present our approach to improving phoneme transition modelling. Building on our previous work, we employ a phoneme partition approach (SME: start, middle, and end states) to build a structure of support vector (SV) classifiers as our main discriminative method. For the phoneme classification step, cross correlation features based on MFCC-vectors are computed and classified within the SME structure. Additionally, we make use of a special reproducing kernel build upon the correlation features, thus offering a direct integration into the SV classifiers. This paper discusses the computation of the afore-mentioned probability outputs as well as initial results using these outputs as emission probabilities in HMMs representing phonemes, applied within a standard speech recognition system.