A novel impulsive noise robust method for mobile telephony speech recognition

Transmission errors commonly encountered in wireless channel usually introduce impulsive noise contamination, which drastically degrades the performance of mobile telephony speech recognition. This paper presents an online method to directly suppress the adverse effect of impulsive noise on speech recognition. In this method, according to noise sensitivity of each feature dimension, the speech observation vector to be recognized is divided into several subvectors, each of which is assigned to a suitable threshold. In recognition stage, observation probability of each feature subvector is floored at the Gaussian mixture level. Thus, the unreliable probability difference caused by impulsive noise is eliminated, and the expected correct path recovers the priority of being chosen in decoding. Experimental results also demonstrate that the proposed method can significantly improve the recognition accuracy both in simulated impulsive noise and machine gun noise environments, while maintaining high performance for clean speech recognition.