ISOLATED WORD SPEECH RECOGNITION EXPERIMENT WITH MICROSOFT WORD FOR AFAAN OROMO TO COMMAND AND CONTROL A COMPUTER SYSTEM

Major Advisor: Jabesa Daba (Das, Professor) Automatic speech recognition (ASR) system is one of the most important technologies that used for human-machine interaction, which is the main goal of ASR. Having similar goals, this study was done to investigate and design an isolated word speech recognition system for Afaan Oromo to command-and-control computer, experiments with Microsoft Word using two most recent open source speech recognition toolkits kaldi and CMUsphinx. To achieve this, literature was reviewed on speech recognitions and Afaan Oromoo speech recognition in particular, human-computer interface, and application of speech recognition, CMUsphinx and kaldi speech recognition. To develop and test the speech input interface required for isolated small vocabulary speaker independent Afaan Oromoo speech recognition to command and control Microsoft Word, sixty(60) Microsoft Word commands were selected, and speech data from 38 people (15 females and 23 males) using the same commands was recorded. To train the recognizer, data recorded from 34(i.e., 89.47% of the recorded audio data) people, and to test the performance of the recognizer, data recorded from 4 people (2 females and 2 males) was used. Once the speech input was developed, a tool was needed to send recognized commands to the Microsoft Windows GUI (in this case Microsoft word) using the recognizer’s output (i.e., text). To do that, pywinauto, a set of Python modules to automate the Microsoft Windows GUI and python programming language was used to make the communication between Microsoft Word and the recognizer. The system performance of the developed system was evaluated using randomly selected 40 commands and the command woordii which was uttered by all speakers to open a Microsoft Word document with a new blank document. The developed system performed 43 commands correctly and five (5) command incorrectly. This experiment show the average overall performance of the system is 89.58 %. Keywords: CMUsphinx, kaldi, Mel-Frequency Cepstral Coefficients (MFCC), acoustic model, Hidden Markov model, Gaussian mixture, SRILM, Monophone training, triphone training