HMM-based Speech Recognition System Research Articles

Hidden Markov models have gained wide acceptance in speech recognition due to the ability to construct optimum (maximum likelihood) models automatically from speech data. Understanding how recognition performance is related to a model's structure is not a simple matter, however, especially where the structure is complex. Thus analysis of errors, especially in terms of the properties of the speech data, is not often undertaken. This paper describes a method for relating speech recognition performance to HMM structure. The basic tool is a best-path (Viterbi) trace of the model through the input data, which is represented using well-known LPC parameters. Linear predictor parameters are used as auxiliary model parameters, independent of the HMM output parameters used for recognition, in order to take advantage of established LPC speech synthesis and LPC speech spectrogram utilities. The trace can then be used to gain insight into error mechanisms, often leading to improvements in model structure and system performance. The description will include techniques for creating the LPC auxiliary model, spectrographic and auditory output from supervised and unsupervised model traces, and an evaluation of the impact of this technique on the development of an HMM-based speech recognition system.