Viterbi Algorithm for multi-pattern joint decoding

Abstract

Multi Pattern Viterbi Algorithm (MPVA) to jointly decode and recognize multiple speech patterns for automatic speech recognition (ASR) is proposed. The MPVA is a generalization of the Viterbi Algorithm (VA) to jointly decode multiple patterns for a given standard Hidden Markov Model (HMM). Unlike our previously proposed Constrained Multi Pattern Viterbi Algorithm (CMPVA), the MPVA does not require the Multi Pattern Dynamic Time Warping (MPDTW) algorithm. The new algorithm has the advantage that it can be extended to connected word recognition (CWR) and continuous speech recognition (CSR) problems. It also gives an improved speech recognition performance over the earlier techniques. Using only two repetitions of noisy speech patterns (-5 dB SNR, 10% burst noise), the word error rate using the proposed MPVA decreases by 28.5 percent, when compared to using individual decoding.