Robust Speaker Recognition with Intra-Speaker Variability Compensation based on Long-Term Recorded Speech Corpus

点击翻译按钮获取中文摘要

This research project aimed to build a new speech corpus that enables many researchers to investigate changes in human voices during a day, a month or several years, and to develop accurate and robust speaker recognition methods for industrial and forensic uses. The speech corpus named "AWA Long-Term Recorded Speech Corpus (AWA-LTR), which is released by Speech Resources Consortium of National Institute of Informatics (NII-SRC), consists of 6 speaker's read speech data recorded at morning, noon, and evening every week for several years (2 to 10 years). Using this corpus, we have developed intra-speaker variability compensation methods that improve the robustness of speaker recognition techniques. We also studied effective speech features for forensic speaker recognition, a comparison between human and machine speaker recognition abilities, accurate and robust speaker modeling methods and speaker verification methods.