最小語者驗證錯誤聯合因素分析

美國國家標準與技術局(NIST)定期於雙數年舉辦語者驗證系統的評比，評比中不但有世界各國研究團隊致力於發展語者驗證系統與其技術交流，同時這項評比也為世界所公認。在過去的NIST SRE08中，聯合因素分析(Joint Factor Analysis, JFA)不但可以處裡通道或環境不匹配所造成的干擾外同時也是僅利用單一系統就能夠抗衡其它團隊提出的複合系統。　　語者驗證系統所做的決策只有目標語者及非目標語者，假使能將這兩種類別的鑑別性拉開，無論是在門檻值的設定或是系統效能的增益都是有所幫助的。而最小驗證錯誤(Minimum Verification Error, MVE)正是這樣的概念，在足夠的訓練資料下利用定義的優化準則逐漸降低量測所得的錯誤並調整語者模型，進而提高目標語者與非目標語者的鑑別性。　　本文基於JFA加上MVE實踐於NIST SRE2010的語料上，在8conv-Core的驗證項目下，外加MVE在最小決策成本能有11.94%的效能增益。

關鍵字

語者驗證；聯合因素分析；最小驗證錯誤； NIST SRE2010

並列摘要

NIST holds a Speaker Recognition Evaluation once every two years; these evaluations not only provide important contribution to the direction of developing system and technical exchange but also are accepted by each research team. In the past of NIST SRE2008, Joint Factor Analysis (JFA) could handle session or channel mismatch, and it could contend with other compound systems at the same time. 　　A speaker verification system just decides target and non-target model. If the system can increases discrimination between the two models, it will help to set the threshold and improve performance. Minimum Verification Error (MVE) is on the same concept; it uses the optimization criterion to decrease mis-verifications and modulates parameter of models when it has enough development corpuses. It produces discriminative speaker model. 　　In this thesis, we propose MVE to speaker verification based on JFA, and it experiments on the NIST SRE2010 corpus. The experimental results on 8conv-Core trial showed that additive MVE furthered 11.94% minimum detection cost value.

並列關鍵字

Speaker verification ； Joint Factor Analysis ； Minimum Verification Error ； NIST SRE2010

參考文獻

[4].D. A. Reynolds, “Channel robust speaker verification via feature mapping,” in IEEE ICASSP, Hong Kong, Apr. 2003, pp. 53-56.

[6].Kenny P., Ouellet P., Dehak N., Gupta V., and Dumouchel P., “A Study of Inter-Speaker Variability in Speaker Verification,” IEEE Transactions on Audio Speech and Language Processing, vol. 16, no. 5, pp. 980-988, Jul. 2008.

[7].Yi-Hsiang Chao, Wei-Ho Tsai, Hsin-Min Wang, and Ruei-Chuan Chang, “Improving the Characterization of the Alternative Hypothesis via Minimum Verification Error Training with Applications to Speaker Verification,” Pattern Recognition, vol. 42, no. 7, pp. 1351-1360, Jul. 2009.

[8].Biing-Hwang Juang, Wu Hou, and Chin-Hui Lee, “Minimum Classification Error Rate Methods for speech Recognition,” IEEE Transactions on Speech and Audio Processing, vol. 5, no. 3, May 1997.

[9].Rong Zheng, Shuwu Zhang, and Bo Xu, “A Comparative Study of Feature and Score Normalization for Speaker Verification,” Lecture Notes in Computer Science, vol. 3832, pp. 531-538, 2005.

國際替代計量

最小語者驗證錯誤聯合因素分析

全文下載

主題瀏覽