Toward more effective acoustic model clustering by more efficient use of data in speech recognition.

摘要

In current large vocabulary continuous speech recognition systems, multivariate Gaussian mixture distributions and context-dependent phones, typically triphones, are used to achieve high accuracy acoustic models. It is crucial to address the problem of how to estimate an extremely large number of model parameters from a limited amount of training data. The traditional approach uses phonetic decision tree based context clustering for reducing free parameters. However, this approach has several problems that might cause system performance degradation. All of these problems are due to the fact that the traditional approach does not efficiently use the limited training data and therefore fails to obtain effective acoustic models. Specifically, three problems are identified and addressed. The first problem is that all states clustered in a leaf node must share the same set of Gaussian components and mixture weights; no distinction is provided among those states. The second problem is due to the fact that triphones that are rarely seen in the training data might be poorly estimated and this causes an adverse effect on decision-tree clustering. The traditional approach lacks an effective mechanism to handle this. The third problem is that only single-Gaussian distributions are used to build decision trees whereas multiple-Gaussian mixture distributions are used in the final model set.; In this thesis, we propose to improve the quality of acoustic models by making use of training data more efficiently. We present a number of ways to address the problems in the traditional approach, namely, a two-level decision tree approach for the first problem, a two-stage decision tree based approach and a MAP-based approach for the second problem, and an approach using a new criterion and an effective clustering algorithm for the third problem. Each of these approaches has successfully reduced the word error rate (WER) of the traditional approach with a statistical significance. Finally the system combining all new approaches has achieved the best performance, which reduced the WER of the baseline system by 14% to 17% relative, with the sizes of acoustic models smaller than those of the baseline models by 8% to 11%.