基於隱藏式條件隨機域模型之千人語者辨識研究

Analysis of HCRF-based modeling for a 1000-speakers identification task

曾家宏(Chia-Hung Tseng)

指導教授 ： 洪維廷

元智大學/電機通訊學院/通訊工程學系/碩士(2012年)

https://doi.org/10.6838/YZU.2012.00244

摘要

本論文旨在研究隱藏式條件隨機域(Hidden Conditional Random Field, HCRF)模型應用於千人語者辨識，並且比較HCRF模型與隱藏式馬可夫模型(Hidden Markov Model, HMM)的效能。經過實驗證實，在訓練以及測試所花費的時間方面，HCRF模型花費少於HMM模型，此外在固定相同系統資源下，HCRF具有更高的辨識率。在訓練模型方面我們提出約束最佳化(Constraint optimization)形式之HCRF訓練法，比傳統HCRF之廣義機率遞減法(Generalized Probabilistic Descent, GPD)訓練方式可得到更低的錯誤率。最後在跨組辨認方面，我們也對HCRF和HMM的表現進行討論。

關鍵字

隱藏式條件隨機域 ； 隱藏式馬可夫模型 ； 語者辨識

並列摘要

In this thesis, we applied the Hidden Conditional Random Fields to a 1000-speakers identification task and compared the performance and computation cost of HCRF with the traditional Hidden Markov Models (HMMs). The experimental results indicate that HCRF models consume less training/testing time than the requirement of HMM; and furthermore HCRFs achieve a higher recognition rate than HMMs with the same system resources. In addition, we propose a constraint optimization method for training HCRF models. The proposed algorithm makes error rate of HCRF model lower than the method using the traditional Generalized Probabilistic Descent (GPD) method. Finally, we also discuss the performance of HCRF and HMM models in cross-group identification.

並列關鍵字

HCRF ； HMM ； speaker identification

參考文獻

[27] W.-T. Hong, “Minimum classification error training of hidden conditional random fields for speech and speaker recognition,” revised in Journal of Information Science and Engineering, 2012.

[2] D.A. Reynolds and R. C. Rose, “Robust text-independent speaker identification using gaussian mixture speaker models,” IEEE Transactions on Speech and Audio Processing, vol. 3, no. 1, pp. 72-83, 1995.

[3] J. Lafferty, A. McCallum, and F. Pereira, “Conditional random fields: probabilistic models for segmenting and labeling sequence data,” in Proceeding of the 18th International Conference on Machine Learning, pp. 282-289, 2001.

[4] W.-T. Hong, “Speaker identification using hidden conditional random field based speaker models,” in Proceedings of International Conference on Machine Learning and Cybernetics, vol. 6, pp. 2811-2816, 2010.

[5] B. H. Juang and S. Katagiri, “Discriminative learning for minimum error classification,” IEEE Transactions on Signal Processing, vol. 40, no. 12, pp. 3043-3054, 1992.