Web-based speech recognition system for Kadazan language

• Main Author: Arfiveyina Armidale
• Format: Academic Exercise
• Language: English; English
• Published: 2022
• Subjects: PL5051-5497 Malayan (Indonesian) languages; QA76.75-76.765 Computer software
• Online Access: https://eprints.ums.edu.my/id/eprint/33269/2/WEB-BASED%20SPEECH%20RECOGNITION%20SYSTEM%20FOR%20KADAZAN%20LANGUAGE.24pages.pdf; https://eprints.ums.edu.my/id/eprint/33269/1/WEB-BASED%20SPEECH%20RECOGNITION%20SYSTEM%20FOR%20KADAZAN%20LANGUAGE.pdf; https://eprints.ums.edu.my/id/eprint/33269/

Because Kadazan speech contains unique traits not seen in other languages, there is currently no system that provides common information and tools for Kadazan speech recognition. In this study, the implementation of Mel Frequency Cepstral Coefficient (MFCC) and Neural Network is explored as a method for the proposed system which is a Web-Based speech Recognition System for Kadazan Language. The objectives of this project are 1) to prepare the requirement and analysis for the Kadazan language speech recognition web-based system, 2) to develop the web-based application for the Kadazan language speech recognition system, and 3) to evaluate Kadazan language speech recognition and functionality of the web-based application. The prototype contains 10 keywords that are used to decide how the users pronounce each of the keywords. The speech recognition technology is incorporated into a web-based system using PHP and Python after extraction, training, and test of the data complete, to create the working implementation that can detect the user's pronunciation accuracy. The output from this project is the system is sometimes unable to predict the words spoken but still gives accuracy. The finding in this project is the MFCC and Neural Network are good feature extraction and classifier. However, to approach the limitation in this project, different feature extraction approaches and the study of additional classifiers, as well as researching by training the model with a larger dataset and using word phonemes are needed.