Compression HMM prototypes (hidden Markov models).

摘要

Compressing HMM prototypes procedure, - be prescribed HMM prototypes (Xj); - reproducing the HMM (Xj) prototypes on compressed HMM prototypes (YJ); - being provided for reproducing the HMM (Xj) prototypes on HMM prototypes tablets (Yj) an encoder (14) which is configured as neural network; - taking the compressed HMM prototypes (Yj) components (YJM, 16), (m = 1, ... M); - being transformed components (YJM) by a bit encoder (22) into binary numbers (Yqjm, 23), (j = 1, ..., J), (M = 1, ..., M); - being provided a neural network (18) for reproducing the compressed HMM prototypes (Yj) on reconstructed HMM prototypes (Xj); - being chosen structure and parameters of the neural network (18) such that the neural network can reproduce the binary numbers (Yqjm, 23) on the HMM prototypes reconstructed (xjk); - being chosen structure and parameters of the neural network (14) of the bit encoder (22) and the neural network (18) such that the binary numbers (Yqjm, 23) are composed of the smallest possible number of bits; and - being chosen neural network (14) and the neural network (18) so that both the distance between (Xj) are minimized and (X''j) as the vector dimension of the compressed HMM prototypes (Yj) .