Confidence-informed unsupervised minimum Bayes risk acoustic model adaptation
Matthew Gibson · 2011
In a supervised speaker adaptation scenario, discriminative techniques have yielded performance gains over standard approaches to adaptation e.g. maximum likelihood linear regression (MLLR). However the discriminative approaches have failed to yield such gains when the task is unsupervised. This paper addresses this issue by applying a novel confidence-informed minimum Bayes risk (MBR) criterion to the task of unsupervised MBR linear regression (MBRLR) speaker adaptation. Experimental evaluations on a large vocabulary recognition task demonstrate that confidence-informed unsupervised MBRLR adaptation delivers significant performance improvements over standard unsupervised MBRLR when using posteriorbased confidence measures. Several aspects of confidence-informed unsupervised MBRLR adaptation are analysed and evaluated, including the use of sub-word confidence measures and MBR criteria, and use of the I-smoothing technique. It is demonstrated that, given improved confidence measures, confidence-informed MBRLR yields performance superior to state-of-the-art confidence-informed MLLR adaptation.