A k-climax neighbors policy based viterbi decoding for large vocabulary mandarin speech recognition

Abstract

In this paper, we apply the k-climax neighbors (k-CN) policy derived from the Bayesian Ying-Yang (BYY) learning framework to Viterbi decoding for Hidden Markov Model based large vocabulary mandarin speech recognition, to adaptively obtain a more precise state decision boundary in the decoding phase. When calculating the posterior probability for each state on a given frame, k Gaussian components from these states are selected by the k-CN policy as the most reliable descriptions, which make the decision boundaries among the competitive candidate states more precise. The experimental results show that a 2.1% relative reduction of the character error rate is achieved on Hub-4 test by adopting the proposed approach.