Frame reliability weighting for robust recognition of partially corrupted speech

A model-based frame reliability weighting method to improve speech recognition when speech signals are partially corrupted by burst noise is proposed. The bias values between speech frames and their corresponding hidden Markov model states are used to represent the reliability of the each frame, serving as the frame weights of a modified Viterbi algorithm. The experimental results show that the proposed frame weighting method effectively represents the importance of each frame and improves the automatic speech recognition performance considerably