Skip to Main Content
Currently, most of the acoustic model selection work is done empirically or heuristically or even arbitrarily. In this paper, Genetic Algorithm (GA) based and Particle Swarm Optimization (PSO) based algorithms that consider the number of states and the kernel numbers for the states simultaneously and reject the uniform allocation of Gaussian kernels are proposed to automatically optimize acoustic model topologies, and some relevant issues are also analyzed and resolved. Experiments on TIDigits corpus show that: first, our GA-based and PSO-based algorithms are effective methods to automatically optimize acoustic model topologies; second, due to the use of Bayesian Information Criterion (BIC), both of our algorithms are capable of achieving higher recognition performance with smaller number of parameters. Specifically, both of our systems with model topologies optimized using GA-based and PSO-based algorithms respectively obtain much increase in recognition performances compared with the baseline systems constructed in a conventional way and having same system complexities; moreover, if compared with baseline systems having same recognition performances, both of our optimized systems save approximate half of the parameters.