A multi-stream speech recognition system based on the estimation of stream weights
Hongyu Guo, Qinghai Chen, Dongmei Huang, Hongyu Guo, Xiaoqun Zhao · 2010 3rd International Congress on Image and Signal Processing · 2010
A multi-stream speech recognition framework based on the estimation of stream weights is proposed for robust speech recognition. First, two complementary acoustic features, MFCCs and LPCCs, were selected. Second, we modeled them separately by using Hidden Markov Models (HMMs), furthermore, formed two streams of this system. Last, we combined the likelihood outputs of the above two systems with weighting technique and obtained a better performance. Here we present a novel algorithm for computing the stream weights of the two feature streams based on the computation of intra-and inter-class distances. Experimental results obtained on Chinese Academy of Science speech database show that this system yields better recognition performance in all conditions. Using this multi-stream framework, we found that the word error rate was decreased by 5%.