Gaussian selection using self-organizing map for automatic speech recognition

The Self-Organizing Map (SOM) is widely applied for data clustering and visualization. In this paper, it is used to cluster Gaussians within the Hidden Markov Model (HMM) of the acoustic model for automatic speech recognition. The distance metric, neuron updating and map initialization of the SOM are adapted for the clustering of Gaussians. The neurons in the resulting map act as Gaussian clusters, which are used for Gaussian selection in the recognition phase to speed up the recognizer. Experimental results show that the recognition accuracy is kept while the decoding time can be reduced by 70%.