Enhanced ASR by acoustic feature filtering

Abstract

Several recent results demonstrate improvement of recognition scores if some FIR ltering is applied on the trajectories of feature vectors. This paper presents a new approach where the characteristics of lters are trained together with the HMM parameters resulting in improvements of the recognition in rst tests.Reestimation formulas for the cut-o frequencies of ideal LPlters are derived as well for the impulse response coe cients of a general FIR LPlter.