Use of Generalised Nonlinearity in Vector Taylor Series Noise Compensation for Robust Speech Recognition

摘要

Designing good normalisation to counter the effect of environmental\uddistortions is one of the major challenges for automatic speech\udrecognition (ASR). The Vector Taylor series (VTS) method is a powerful\udand mathematically well principled technique that can be applied\udto both the feature and model domains to compensate for both\udadditive and convolutional noises. One of the limitations of this\udapproach, however, is that it is tied to MFCC (and log-filterbank)\udfeatures and does not extend to other representations such as PLP,\udPNCC and phase-based front-ends that use power transformation\udrather than log compression. This paper aims at broadening the\udscope of the VTS method by deriving a new formulation that assumes\uda power transformation is used as the non-linearity during\udfeature extraction. It is shown that the conventional VTS, in the log\uddomain, is a special case of the new extended framework. In addition,\udthe new formulation introduces one more degree of freedom\udwhich makes it possible to tune the algorithm to better fit the data\udto the statistical requirements of the ASR back-end. Compared with\udMFCC and conventional VTS, the proposed approach provides upto\ud12.2% and 2.0% absolute performance improvements on average, in\udAurora-4 tasks, respectively