A compensation method for recognition of speech under G-force

A compensation factor is proposed, which can describe the relations between the neutral speech and the speech under stress. And the stress compensation factor also considers both the shift of means and the differences of the variances caused by the stress. Based on the parameters initialized by k-means method, the final results of the compensation factor are estimated by the EM algorithm. Through the compensation factor, the performance of a recognition system can be enhanced. Under the G-force stress, the proposed method can get 32.3% improvement for a small-vocabulary speaker-dependent system.