Robust speaker adaptation based on parallel factor analysis of training models

The two major discrepancies between the training and deployment conditions in automatic speech recognition are speaker and noise environment. Presented is a speaker adaptation method which is robust to noise environments in the framework of the basis-based technique. A training tensor composed of speaker-dependent models is decomposed by parallel factor analysis, which can produce the bases that are more robust and compact than those obtained by principal component analysis. Experimental results show that the proposed method performed as good as the eigenvoice in a clean environment and outperformed the eigenvoice in noise environments.