Spectrum filtering with FRM for robust speech recognition

When a speech recognition system is used in real environment, noises have a serious problem in the system. In this paper, we propose a method which normalizes feature parameters after filtering the time trajectories of both the power spectra and the log power spectra. In order to improve the recognition accuracy, the designed FIR filter become considerably high order. Using a FRM filter, we can develop the robust speech recognition whose recognition ability is high enough and whose filtering order can be reduced low. As a result of the isolated word recognition experiment where fifteen kinds of noises were added, higher recognition rates was obtained than conventional method by spectrum subtraction (SS) and relative spectral (RASTA)