Constructing noise robust speaker verification system based on kernel method

2014 Fiscal Year Final Research Report

• Project/Area Number: 25880026
• Research Category: Grant-in-Aid for Research Activity Start-up
• Allocation Type: Single-year Grants
• Research Field: Intelligent informatics
• Research Institution: Tokyo Metropolitan University (2014); The Institute of Statistical Mathematics (2013)
• Principal Investigator: SHIOTA Sayaka 首都大学東京, システムデザイン研究科, 助教 (90705039)
• Research Collaborator: MATSUI Tomoko ; OGAWA Tetsuji ; KIYA Hitoshi ; MARKOV Konstantin ; GRETTON Arthur ; PETERS Gareth
• Project Period (FY): 2013-08-30 – 2015-03-31
• Keywords: 話者照合 / 音声信号処理 / 統計的機械学習 / カーネル法

Outline of Final Research Achievements

Speaker verification system is one method of biometrics authentication system. The conventional speaker verification system based on statistical machine learning methods require enough number of training data for training speaker dependent models. In this research, a kernel based method is used for calculate two samples distance. The distance is called MMD (Maximum Mean Discrepancy) and the two sample test is based resampling method. However, I found that the MMD based resampling method is not suitable for time series data. Recently, a wild bootstrap technique based MMD method has been reported for time series data. Thus, the wild bootstrap based MMD method is carried out for speaker verification system. Since the proposed method uses a kernel framework, some parameters are affected sensitively but a small experiment achieved a high accuracy score. In future work, the some conditions should be tested.

Free Research Field

音声信号処理