Forward/Backward Algorithms for joint multi pattern speech recognition

We are addressing the problem of jointly using multiple noisy speech patterns for automatic speech recognition (ASR), given that they come from the same class. If the user utters a word K times, the ASR system should try to use the information content in all the K utterances of the word simultaneously and improve its speech recognition accuracy compared to that of the single pattern speech recognition. In this paper we propose two types of Multi-pattern Joint Likelihood Forward Backward (MJLFB) algorithm to address this problem. We show an analysis of the differences between the two types of MJLFB algorithms. We also propose a new criteria (which does not require any threshold) to calculate joint probability of the feature vectors emitted given the Hidden Markov Model (HMM) state. The new formulation is tested in the context of speaker independent isolated word recognition (IWR). When 10 percent of speech is affected by burst noise at -5 dB SNR (local), it is shown that joint recognition using only two noisy speech utterances reduces the noisy speech recognition error rate by 52.40 percent, when compared to the single pattern recognition using the Forward Algorithm (FA).