Time and Frequency Filtering for Speech Recognition in Real Noise Conditions

MFCCs perform well when used for clean speech recognition. However, for noisy speech the recognition rates go down. Augmenting the MFCC feature vector by dynamic features improves both discrimination and robustness of the MFCC-based recognizer. In this paper, we present an alternative para meterization based on the frequency filtering (FF) technique. By using FF, a significant improvement with even lower computational costs can be obtained for both clean and noisy speech recognition rates in comparison to MFCC (especially for flat-spectrum noises). An additional improvement can be achieved by using features obtained through properly designed time filters instead of the usual delta and acceleration features. In a few preliminary tests with the Aurora database in the framework of the performance evaluation of DSR front-ends, 25,8% and 9,8% relative improvements were achieved for clean and noisy speech recognition, respectively, only by using typical time and frequency filters instead of the reference parameterization.