Speaker adaptation of output probabilities and state duration distributions for speech recognition

This paper presents a comparison of maximum a posteriori (MAP) speaker adaptation of state duration distributions and output probabilities in HMM. Both adaptation procedures are compared and then combined in recognition experiments with clean and noisy signals. The results here shown suggest that the state duration distribution adaptation can lead to higher improvements than the adaptation of output probabilities, and the reduction in the error rate when both adaptations are combined is as high as 50% or 60% using only a few samples per word.