syatalio Ned Network 4rOhiteotnre for Seoond Order Hidden Markov Model8
Feng Zhaozhi Huang, Zailu Chen, P. R China · 1992
This paper presents a systolic neural network architecture for implementing second order hidden Morkov models (SOHMM's!. A programmable systolic arrays is proposed. A unified model for recurrent high order multilayer feedforward neural networks and SOHMM's is exploited for the architecture design. Extended Viterbi algorithm for SOHMM's is described. Final lyr the implementation based on TNS320C25 chip is also discussed. 1: Introdaetlon With the development of %SI, a wide surge of interest has been brought into the research area of HMM's and MLP's or the hybrid algorithm of both methods. However, all previous works are only limited to HLP's and first order HMH. In order to improve the recognition accorate of speech, the authors have proposed a new high order multilayer feedforward neural network (HM), and pointed out that the second order HNMSOHHH) can be viewed as a special case of recurrent HH (RHM)ClIC21. The amazing algorithmic analogy between HM and SOHMM will lead to a unified mathematical formulation for both models. This paper is arranged as following. In section 2 a unified algorithmic formalation for RHNNN's and SOHMM's is exploited for the systolic architecture design. Section 3 presents the systolic design for the retrieving and learning phases of the unified model. The extended Viterbi algorithm for second order HMM's is described in section 4.