Bayesian Learning for Deep Neural Network Adaptation

A key task for speech recognition systems is to reduce the mismatch between training and evaluation data that often attributable speaker differences. Speaker adaptation techniques play a vital role mismatch. Model-based approaches require sufficient amounts of target ensure robustness. When amount level limited, prone overfitting poor generalization. To address issue, this paper proposes full Bayesian learning based DNN framework model speaker-dependent (SD) parameter uncertainty given limited specific data. This investigated in three forms techniques: hidden unit contributions (BLHUC), parameterized activation functions (BPAct), bias vectors (BHUB). In methods, deterministic SD parameters are replaced by latent variable posterior distributions each speaker, whose efficiently estimated using variational inference approach. Experiments conducted on 300-hour speed perturbed Switchboard corpus trained LF-MMI TDNN/CNN-TDNN suggest proposed consistently outperform NIST Hub5'00 RT03 sets. only first five utterances from as data, significant word error rate reductions up 1.4% absolute (7.2% relative) were obtained CallHome subset. The efficacy further demonstrated comparison against state-of-the-art performance same most recent reported literature.