| Most research and development in current speech recognition technology uses statistical modeling, or in particular, a hidden Markov model (HMM). The versatility of the HMM stems from the fact that it can learn statistical properties of the speech inductively from pre-classified training examples. The maximum likelihood (ML) criterion is typically used and an expectation maximization (EM) algorithm can be employed to determine the parameters from the expectation of the underlying process. There are a large number of parameters in an HMM and, as a consequence, a large amount of data is required in order to guarantee enough training tokens for each parameter. The HMM training process is simple, well-defined, and numerically stable, but the large amount of training data often causes the process to converge rather slowly. Many different techniques have been suggested that typically trade-off between computational complexity and performance.;The main goal of this thesis is to achieve an efficient training of an HMM using an incremental variant of the EM algorithm. The algorithm estimates the model parameters from a subset of the training data, and then iterates until convergence. The training strategy contrasts sharply to the standard batch training where the model is updated only after all the data are processed. The ML criterion can be used, however, maximum a posteriori (MAP) estimation plays an important role when incremental training is in progress. The latter combines the information known a priori and an evidence observed from the training data. An appropriately chosen prior pdf allows the EM training steps to go in right direction with-out enforcing too much restriction. Experimental results show that the approach converges substantially faster without any loss of recognition performance. Although the discussion is primarily aimed at improvement of training efficiency, the work does have greater implications. Several related topics, including theoretical development and simulation studies, are presented in order to investigate the features of tools from various perspectives. One notable achievement is that the same approach can be used to adjust a baseline model to a new acoustic environment (a remote, microphone-array system) with minimal degradation. |
