| Abstract [eng] |
The aim of this project is to investigate the recognition accuracy of Lithuanian names and digit names therefore to create a demo version of disease codes recognizer interface. The object of the project – disease codes which are based on the recognition of Lithuanian names and digit names. The goals of the research: 1. Perform an analysis of literature on speech recognition related topics. 2. Prepare the non-native language recognizer and investigate its recognition accuracy. 3. Master the operating principles of HTK suite and investigate the accuracy of Lithuanian recognizer. 4. Build a demo version of the recognizing software. 5. Provide conclusions and results. The following methods used in this research: descriptive, experimental, quantitative and comparison method. In studies of the non-native language recognizer, the dependence of the speaker profile and its training were being investigated. The diferences in recognition accuracy between results of the synchronous and asynchronous modes were found. Using the HTK suite a new Lituanian recognizer was built to obtain and compare the results by changing the number of states and in addition to varying the number of Gaussian mixtures. The test results revealed that there is no significant change in recognition whether using trained or not trained user profile. All four test profiles provided results which vary from 94,12% to 96,44%. Although it turned aut the the asynchronous mode is capable of providing better results than the synchronous mode – it is 93,7% compared to 89,12% provided by the synchronous mode test results. The analysis of the Lithuanian speech recognizer was conucted in two phases. |
