Robust speech recognition with feature extraction using combined method of RSF and DRA

The paper explores the extraction of speech features aiming at noise robustness for speech recognition and proposes advanced speech analysis techniques named RSF/DRA (running spectrum filtering/dynamic range adjustment). The proposed techniques, DRA and RSF, focus on speech feature adjustment. DRA normalizes cepstral dynamic ranges and RSF eliminates the jitter influences of speech feature parameters. Experiments on isolated word recognition were carried out using 40 male and 40 female speakers for training and 5 male and 5 female speakers for recognition. The results of the recognition rate improving from 17% to 63% versus running car noise at -10 dB SNR show the effectiveness and high noise robustness of the proposed method.