(A) study on the performance improvement of thespeech recognition system based on the phoneme-level hidden markov model

Abstract

The need for a speaker-independent large vocabulary speech recognition system has been grown due to its large application area. Although the phoneme-level HMM has been widely used as an efficient algorithm for large vocabulary, the performance of the phoneme-level HMM-based recognition system have to be improved more form practical use. For this reason, a new HMM parameter estimation algorithm and a VQ codebook design procedure are proposed to raise recognition accuracy, and an efficient pre-classification algorithm is proposed to reduce recognition time. In order to establish a benchmark performance, a phoneme-level HMM-based recognition system is first implemented as a baseline system. And then, the performance of the baseline system is improved. First, two methods of improving HMM parameter estimation algorithm are presented. The first one is an HMM parameter estimation method minimizing error rate. In this method, a performance function which is proportional to the accuracy of the recognition system is introduced. HMM parameters are estimated so that the performance function can be maximized. The proposed algorithm emphasizes training patterns making errors or near-misses. Applying this algorithm to the baseline systems, the error rate for training data is significantly reduced, but the error rate for test data is slightly decreased. The second method is an HMM parameter smoothing method based on the fuzzy mapping concept. In this method, HMM parameters are smoothed by the smoothing matrix obtained by the fuzzy relationship between output symbols and training data. The fuzzy smoothing method reduces the error rate of the baseline system by approximately 50 percent. Second, a VQ codebook design algorithm integrated with HMM is proposed so that the discrimination ability and the robustness of HMM parameters can be improved. For this purpose, we extract codewords from the state segments of each recognition unit by an MKM algorithm or an LVQ2 algorithm, where t...