Efficient computation of the frame-based extended union model and its application in speech recognition against partial temporal corruptions

摘要

The extended union model (EUM) was recently proposed and shown to be effective in handling short time temporal corruption. Because of the computational complexity, the EUM probability can only be computed over groups of consecutive observations (called segments) and recognition can only be performed under N-best re-scoring paradigm. In this paper, we introduce a hidden variable called “pattern of corruption” and re-formulate the extended union model as marginalizing over possible patterns of corruption with likelihood computed via the missing feature theory. We then introduce a recursive relationship between the EUM probabilities of two successive observation sequences that can greatly simplify the EUM probability computation. This makes it possible to compute the EUM probability over a long sequence. Using this recursive relationship, the EUM probability over frames, called the “frame-based EUM” can easily be computed. To simplify the EUM-based recognition, we propose an approximated, dynamic programming-based EUM recognition algorithm, called the Frame-based EUM Viterbi algorithm (FEVA), that performs recognition directly instead of via N-best re-scoring. Experimental results on digit recognition under added impulsive noises show that both the frame-base EUM and the FEVA outperform the segment-based EUM.