next up previous contents
Next: Limitations and gains Up: Description of the model Previous: Output density models.

HMM/ANN hybrids.

The rapid development both in the computer hardware and in the neural computation methods has made it possible to apply many attractive density models for the HMMs to constitute hybrid HMM/ANN recognition systems [Lippmann, 1989,Morgan and Bourlard, 1995]. The densities can be approximated by, for example, multilayer perceptrons (MLPs) [Bourlard and Wellekens, 1990,Franzini et al., 1990,Renals et al., 1994], time delay neural networks (TDNNs) [Dugast et al., 1994] or radial basis function networks (RBFs) [Huang and Lippmann, 1991,Singer and Lippmann, 1992,Renals et al., 1994]. The Gaussian RBFs are actually very close to Gaussian kernel density estimators; the difference is only in the way they are trained [Renals et al., 1991,Renals et al., 1992]. The ANNs can also be used for subtasks in the density modeling, as for generating the VQ codebook by LVQ [Iwamida et al., 1990,Torkkola et al., 1991,Makino et al., 1992] or smoothing the VQ weights parameters by SOM [Zhao and Rowden, 1991,Monte, 1992,Kim et al., 1994].

In this thesis the ANNs (SOM and LVQ) are applied to assist in the training of the MDHMMs to develop methods that provide low error rates with a tractable amount of computations. The applications to the density initialization is discussed in Publications 1 and 3, to the actual training of the HMMs in Publications 4 and 5 as an enhancement of the conventional Viterbi training, and to the corrective tuning in Publication 2. Publication 6 compares then the results of some options for the total training process. Some of the main alternative HMM training methods have been presented, for example in [Kurimo, 1994] and [Morgan and Bourlard, 1995].


next up previous contents
Next: Limitations and gains Up: Description of the model Previous: Output density models.
Mikko Kurimo
11/7/1997