摘要： 针对说话人辨认中训练语音有限时系统泛化能力差的问题，提出一种基于半监督学习的复合高斯混合模型算法。通过复合高斯混合模型对所有说话人的特征分布统一建模，基于半监督学习机制下的EM算法对学习样本进行学习。实验证明，该算法能够充分利用未标记样本对系统进行有效的自适应更新，改善系统的性能，获得比传统高斯混合模型更高的识别率，提高系统的泛化能力。

关键词: 说话人辨认, 半监督学习, 复合高斯混合模型, EM算法

Abstract: Aiming at the limitation of generalization ability when the speaker identification system falls short of training data, this paper proposes a new method of complex Gaussian Mixture Model(GMM) based on Semi-Supervised Learning(SSL). It uses complex GMM to model the feature distribution of all speakers and the training period adopted SSL based on iterative algorithm. Experimental shows that this new method can take full advantage of unlabeled data on the self-adaptive updating, improve the performance, get higher recognition rate compared with traditional GMM, and effectively enhance the generalization ability of system.

Key words: speaker identification, Semi-Supervised Learning(SSL), complex Gaussian Mixture Model(GMM), EM algorithm

中图分类号: 

• TP18