### Two variants in HMM ASR with soft missing data - outline of theory and initial
experiments

#### Andrew Morris

#### IDIAP, Switzerland

#### 1. Combining FC experts, MD theory and EM for ML expert weights estimation

We first recap on the EM equations for mix weight estimation in the simple case when individual mix component pdfs are fixed. We then show how these simple equations can be adapted for the purpose of offline estimation of subband expert weights, in the case where both component pdfs and mix weights are fixed, but a separate model is been trained for every subband combination, and the likelihoods from these models are to be combined in a linearly weighted sum.

#### 2. ASR with soft missing data, where uncertainty is based on data utility as well as reliability

After briefly revisting the original equations underlying the marginal/bounds decomposition normally used in recognition with missing data, it is shown how these equations can be used to provide a model for softening the observed/missing decision. This model is a generalisation of a recently proposed decision softening approach which significantly increases recognition performance with missing data.

Jon Barker
Last modified: Tue Jan 30 11:37:48 GMT 2001