On Separating Environmental and Speaker Adaptation

Abstract

This paper presents a maximum likelihood (ML) approach, concerned to the background model estimation, in noisy acoustic non-stationary environments. The external noise source is characterised by a time constant convolutional and a time varying additive components. The HMM composition technique, provides a mechanism for integrating parametric models of acoustic background with the signal model, so that noise compensation is tightly coupled with the background model estimation. However, the existing continuous adaptation algorithms usually do not take advantage of this approach, being essentially based on the MLLR algorithm. Consequently, a model for environmental mismatch is not available and, even under constrained conditions a significant number of model parameters have to be updated. From a theoretical point of view only the noise model parameters need to be updated, being the clean speech ones unchanged by the environment. So, it can be advantageous to have a model for environmental mismatch. Additionally separating the additive and convolutional components means a separation between the environmental mismatch and speaker mismatch when the channel does not change for long periods. This approach was followed in the development of the algorithm proposed in this paper. One drawback sometimes attributed to the continuous adaptation approach is that recognition failures originate poor background estimates. This paper also proposes a MAP-like method to deal with this situation.