Running spectrum filter for robust speech recognition

摘要

When a speech recognition system is used in real environment, noises cause a serious problem in the system. In this report, we propose the method for decreasing the noise influences with a running spectrum filter. The running spectrum is represented as the time trajectories of spectra. It is a very important characteristic in speech recognition. The proposal method makes the low-pass filter apply on the time trajectories of the power spectrum and the band-pass filter apply on the time trajectories of I lie log power spectrum. As a result of the isolated word recognition experiment where four kinds of noises were added artificially, higher recognition rates was obtained especially in lower SNR environment compared with RASTA and CMS.