An Isolated Word Recognizer for the Swahili Dialect Using Sphinx4-Neural Network Hybrid

Speech is one of the most effective means of communication. Efforts to develop an automatic speech recognizer for Swahili dialect has faced numerous challenges due to lack of a standardized acoustic model for the dialect. This study aimed at developing an isolated speech recognizer for automatic transcription of Swahili dialect using Sphinx4 Neural Network hybrid. The objectives of the study was to design and develop a neural network on the front-end module of the Sphinx4 and to compare performance between sphinx4 and sphinx4 neural network hybrid. The study was guided by statistical and probabilistic theory and the theory of linguistics. This study followed a positivist philosophical paradigm and embraced experimental research design. To achieve this, the researcher developed a small corpus of the dialect which contained twenty words. A total of two hundred speech recordings were made from ten volunteers. The words and the volunteers were selected using purposive and convenience sampling methods respectively. Results obtained from both sphinx4 and sphinx4-neural network hybrid were evaluated using descriptive analysis techniques. This study established that a neural network could be integrated with sphinx4. The study also established that sphinx4-neural network hybrid with its neural network trained to an error less than 0.0175 had a performance that was statistically at par with that of sphinx4. The study concluded having established that for isolated word recognition in Swahili, sphinx4-neural network hybrid can be used in place of sphinx4 HMM recognizer. Areas of further study recommended include; using this tool for continuous speech recognition, evaluation of the tool using a larger speech sample, comparative analysis between the tool and others line NICO and establishing ways of improving accuracy of the tool.