Neural Networks For Speech Recognition Of A Phonetic Language

The goal of this thesis is to explore a possibility for a viable alternative/replacement to the Amharic typewriter. Amharic is the national language of Ethiopia. It is one of the oldest languages in the world. Actually, the root-language of Amharic, called Geez, is a descendent of Sabean, which is the direct ancestor of all Semitic languages including English. A phonetic language with 276 phonemes/characters, Amharic has posed quite a challenge to those who,like the author of this thesis, have attempted to design an easy-to-use word processor that interfaces with the conventional keyboard. With current Amharic word processing software, each character requires an average of threekeystrokes thus making typing Amharic literature quite a task. This thesis researches the feasibility of developing a PC-based speech recognition system to recognize the spoken phonemes of the Amharic language. Artificial NeuralNetworks are used for the recognition of spoken alphabets that form Amharic words. A neural network with feed-forward architecture is trained with a series of alphabets and is evaluated on its ability to recognize subsequent testdata. The neural network used in this project is a static classification network; that is, it focuses on the frequency domain of speech while making no attempt to process temporal information. The network training procedure uses thegeneralized Delta Rule. The recognition system developed in this project is an Isolated Speech Recognition System. The approach taken is to recognize the spoken word character by character. This approach is expected to work well dueto the phonetic nature of Amharic.