A SOM based 2500 - Isolated - Farsi - Word speech recognizer

A modified 2-D Kohonen Self-Organizing (MSOM) neural network is used for recognizing Farsi isolated words. The network dimension is 10*15 cells with a hexagonal topology and it is trained using 300 Farsi words. As input vectors for learning, speech spectrum and energy of signal are used. The weight vectors of the cells are then fine tuned using supervised learning vector quantization 3 (LVQ3) technique. The cells are labeled to 28 out of 29 Farsi phonemes. At the word recognition stage, the quasi phonemes are obtained. Then the phonemes are determined. Using the phonetic rules of Farsi words and the connection rules of Farsi characters, the recognized word will appear on the monitor. To remedy the errors, a 2500 word dictionary is used. The determined sequence of phonemes is given to the dictionary, and the closest word to the sequence is shown on the monitor. The proposed recognizer is able to recognize all vowels with the accuracy of 100 percent, and it also recognize correctly 55 isolated words among 100 words.