Weighted parallel model combination for noisy speech recognition

This paper proposes a modified parameter mapping scheme for parallel model combination (PMC) method. The modification aims to improve the discriminative capabilities of the compensated models. It is achieved by the rearrangement of the distributions of state models in order to emphasize the contribution of the mean in the following process. Both distributions of speech model and noise model are shaped in cepstral domain through a covariance contracting procedure. After the compensation steps, an expanding procedure of the adapted covariance is necessary to release the emphasis. Using this process, the discriminative capability is increased so that the recognition accuracy is improved. In this paper, the recognition of Chinese names demonstrates the improvement to the original PMC method, especially when SNR is low.