Mobile phone speaker recognition

• 主要作者: Thang, Hui Ru.
• 其他作者: Chng Eng Siong
• 格式: Final Year Project
• 語言: English
• 出版: 2013
• 主題: DRNTU::Engineering::Computer science and engineering
• 在線閱讀: http://hdl.handle.net/10356/55031
• 機構: Nanyang Technological University

Today in a society with well-developed technology, a high penetration rate of mobile devices clearly illustrated how well-woven our portable phones are, into our lives. Therefore, security and privacy is a growing issue. To solve this, biometric systems such as speaker recognition can be used to increase protection level and security. The purpose of this project is to implement a speaker recognition system in Android platform, on S3. The system should allow users to record his speech utterance, extract MFCC features and test it against the speaker’s model. If the utterance is from the speaker, the system should accept it and reject if otherwise. Server-client architecture is used, where UBM is trained at server side and preloaded into android devices (client) in order to perform the verification task. Experimental results showed peak performance when 256 GMM was used. Above that, the problem of over-fitting occurs. More training data can be supplied and experiments can be extended to out-of set testing. Future enhancement includes storing the UBM online in a database to avoid manual transfer of UBM when large number of clients is used. The current verification system can also act as a baseline system and be compared with other systems created in the future, such as iVector and GMM-SVM.