论文信息 - A Feature Space Transformation Method for Personalization using Generalized I-Vector Clustering

A Feature Space Transformation Method for Personalization using Generalized I-Vector Clustering

We present a feature space transformation method for personalization. This method includes a generalization of i-vector based clustering that allows parameter tying of sub-loading matrices. This method trains i-vector parameters from the utterances of a mobile device, uncovering a low dimension space for clustering variability given the device. We show through empirical results impact of parameters of the generalized i-vector method. We conducted recognition experiments on an internal large vocabulary voice search system for gaming. The method achieved significant reductions of word error rates by 28%, compared to a per utterance adaptation system.

Kaisheng Yao | Yifan Gong | Chaojun Liu

[1] Zhi-Jie Yan,et al. An i-vector Based Approach to Training Data Clustering for Improved Speech Recognition , 2011, INTERSPEECH.

[2] Brian Kingsbury,et al. Boosted MMI for model and feature-space discriminative training , 2008, 2008 IEEE International Conference on Acoustics, Speech and Signal Processing.

[3] Martin Karafi. iVector-Based Discriminative Adaptation for Automatic Speech Recognition , 2011 .

[4] Mark J. F. Gales,et al. Maximum likelihood linear transformations for HMM-based speech recognition , 1998, Comput. Speech Lang..

[5] Daniel Povey,et al. The Kaldi Speech Recognition Toolkit , 2011 .

[6] Zhi-Jie Yan,et al. A new i-vector approach and its application to irrelevant variability normalization based acoustic model training , 2011, 2011 IEEE International Workshop on Machine Learning for Signal Processing.

[7] Kaisheng Yao,et al. A basis representation of constrained MLLR transforms for robust adaptation , 2012, Comput. Speech Lang..