Unsupervised Acoustic Model Adaptation Based on Ensemble Methods
Takahiro Shinozaki, Yu Kubota, Sadaoki Furui · IEEE Journal of Selected Topics in Signal Processing · 2010
We propose unsupervised cross-validation (CV) and aggregated (Ag) adaptation algorithms that integrate the ideas of ensemble methods, such as CV and bagging, in the iterative unsupervised batch-mode adaptation framework. These algorithms are used to reduce overtraining problems and to improve speech recognition performance. The algorithms are constructed on top of a general parameter estimation technique such as the maximum-likelihood linear regression method. The proposed algorithms are also useful for suppressing the negative effects of unsupervised adaptation, which reinforces the errors included in the hypothesis used for the adaptation. Experiments are performed using clean and noisy speech recognition tasks with several conditions. We show that both our proposed unsupervised adaptation algorithms give higher performance than the conventional batch-mode adaptation algorithm; however, the unsupervised CV adaptation algorithm is more advantageous than the unsupervised Ag adaptation algorithm in terms of computational cost. The proposed algorithms resulted in 4% to 10% relative reduction in the word error rate over the conventional batch-mode adaptation.