New continuous speech feature adjustment for a noise-robust CSR system

Abstract

We propose a noise-robust continuous speech recognition (CSR) method for recognition. In model building, we extract the novel feature vector by using running spectrum analysis (RSA) and dynamic range adjustment (DRA) methods. DRA adjusts the dynamic range on MFCC modulation spectrum domain (MSD). In recognition, the algorithm automatically divides the continuous speech into short sentences and blocks, then we use DRA based on the blocks. The proposed algorithm efficiency is studied for clean and noisy environment. In our experiments, all HMMs have been trained by using the Japanese newspaper article sentence (JNAS) database. The average recognition rate improves under various types of noise and SNR conditions.