Energy-based Competitive Learning Algorithm And Its Application

The goal of speech recognition is able to make the machine to understand human language. In various circumstances, the machine can accurately identify the contents of the speech and execute the command according what it has identified.Vector quantization is one of the best techniques with good performance of the model training and pattern matching in speech recognition technology. As we all know, the codebook design is the primary and core issues of the vector quantization. The optimal codebook is the premise of the implementation of vector quantization techniques, so the importance is very important for the design of the codebook. Classic LBG algorithm has an advantage of fast convergence, although can easily fall into local optimum. The initial codebook has a great impact on the codebook design. The algorithm requires the size of codebook which must be given as a known condition. This can’t be easy to make in real-world conditions.In the past few decades, the competitive learning has received wide concern. Due to its adaptive online learning, competitive learning has been widely applied to data clustering and some other fields. Energy based competitive learning algorithm which has been came up by Chang-dong Wang and Jian-huang Lai has some advantages which is listed as follows.(1) It can be initialized automatically, which means that the algorithm is able to determine the appropriate number of clusters and initial codebook automatically.(2) Different seed points have different adaptive learning rate. For a sample point, the seed of the winning cluster will be upgraded with different learning rate according the energy of the sample.(3) The algorithm can eliminate the impact of the overlapping points among clusters or the isolated points on the clustering results. Learning rate is based on energy of samples and the energy of overlapping points and isolated points is small. So these points almost have no effect on the update of the seed points.In view of the above-mentioned advantages of the energy based competitive learning algorithm in data clustering, this algorithm has been applied to digital speech recognition of specific person to verify the advantage of the algorithm. The simulation results show that the speech recognition based on energy-based competitive learning algorithm is better than LBG algorithm.