POSSIBILITY ASSESSMENT OF USING TIGRIGNA SPEECH TO COMMAND AND CONTROL A COMPUTER APPLICATION

Abstract

This study is aimed at exploring the possibility of developing Tigrigna speech input interface to command and control Microsoft Word. The prototype Tigrigna speech input interface developed in this experiment has two subsequent systems namely speech command recognition and command execution. In the course of developing the speech recognition system, 50 command words used to command and control Microsoft Word were randomly selected, translated to Tigrigna and isolated utterances of these words were captured from 40 speakers (20 female and 20 male) in the age range of 20-50. Those speakers were from three different Tigrigna accents (Raya, Meqele and Northern & Central Tigray (NCT)). Then, 75% of the recorded dataset was used to train and build the recognizer and the remaining 25% was used to test system’s performance. To assess the possibility with the available data two techniques were employed. First, considering vocabulary size as a factor that can determine system’s performance, five HMM based isolated Tigrigna word recognizers with vocabulary size of 10, 20, 30, 40 and 50 command words were trained and evaluated. Second, besides vocabulary size, considering accent variation as a factor, a model with vocabulary size of 20 words was trained and evaluated for NCT accent only; where the list of the command words used in this model was the same as these used by the second model from the first five models. After all performance of these recognizers were analyzed using the prepared test dataset and evaluation results were 98.98%, 98.76%, 99.17%, 99.38%, 99.5% and 99.75% respectively. Furthermore, the sixth recognizer was analyzed using live audio as it was further considered to develop the speech input interface. The test speech was taken from six (3 male and 3 female) speakers and the system performance was 88.33%. Finally, the sixth recognizer was integrated with C# and Tigrigna speech input interface was designed and developed for Microsoft Word. To ensure that the whole system is working properly a test was conducted using live audio taken from one speaker; and experimental results has indicated that, despite the accuracy issues associated with the recognizer, the speech input interface developed was able to execute what has been recognized correctly. Overall, despite all the factors, experimental results has clearly indicated the feasibility of developing isolated Tigrigna word recognizer as well Tigrigna speech input interface to command and control Microsoft Word.