Channel compensation for robust telephone speech recognition

Channel compensation has been proved to be an effective approach for robust speech recognition. We first use two level cepstral mean substraction (CMS) for channel compensation in telephone speech recognition, and then propose a new channel compensation method, RMFCC. For all experiments, a Korean isolated 84 words database consisting of 80 speakers collected from a local telephone line is adopted. Using the RMFCC, a 39.8% reduction in word error rate is obtained relative to the conventional HMM system. It is shown from the experiments that the RMFCC, compared with RASTA, reduces the computational complexity without losing accuracy, and is also better than the CMS and two level CMS in performance.