Self-organizing Maps for Speech Recognition

The spoken speech is the easiest and most natural way for the communication between human beings. So, the human-machine communication can be executed based on the way that human-human communication occurs. Researches in automatic speech recognition (ASR) have been developed for decades to produce communication as natural as possible. There some few attempts to use Self-organizing Maps to solve ASR problems, often working to execute pattern recognition. In this paper, we comparatively analyze the efficiency of two different neural networks, the Self-Organizing Maps (SOM) and the Time Organized Maps (TOM), applied for the recognition of the American English phonemes. We considered phonological features to represent the input data. The results of the experiments suggest that the SOM is more efficient than TOM, even with simulations of disturbed data, including noise that may appear and harm the input signal quality. Keywords— Speech Recognition; Phonemes Recognition; SelfOrganizing Maps; Time-Organized Maps; Phonological Features.