Issues with uncertainty decoding for noise robust automatic speech recognition

Abstract

Interest continues in a class of robustness algorithms for speech recognition that exploit the notion of uncertainty introduced by environmental noise. These techniques share the property that the uncertainty varies with the noise level and is propagated to the decoding stage, resulting in increased model variances. In observation uncertainty forms, the uncertainty variance is simply the variance of the error in enhancement that is added to the model variances. Another form, called uncertainty decoding, refers to a factorisation which results in a linear feature transform and model variance bias that increases with noise; using appropriate approximations, efficient implementations may be obtained, with the goal of achieving near model-based performance without the associated computational cost. Unfortunately, uncertainty decoding forms that compute the uncertainty in the front-end and pass this to the decoder may suffer from a theoretical problem in low signal-to-noise ratio conditions. This report discusses how this fundamental issue arises, and demonstrates it through two schemes: SPLICE with uncertainty and front-end joint uncertainty decoding ( FE-Joint). A method to mitigate this for FE-Joint compensation is presented, as well as how SPLICE implicitly addresses it. However, it is shown that a model-based joint uncertainty decoding approach does not suffer from this limitation, like these front-end forms do, and is more computationally attractive. The issues described and performance of the various schemes are examined on two artificially corrupted corpora: the AURORA 2.0 digit string recognition and 1000-word Resource Management tasks.