Bayesian Learning of LF-MMI Trained Time Delay Neural Networks for Speech Recognition

Discriminative training techniques define state-of-the-art performance for automatic speech recognition systems. However, they are inherently prone to overfitting, leading poor generalization when using limited data. In order address this issue, paper presents a full Bayesian framework account model uncertainty in sequence discriminative of factored TDNN acoustic models. Several learning based variant systems proposed the over weight parameters and choices hidden activation functions, or layer outputs. Efficient variational inference approaches as few one single parameter sample ensure their computational cost both evaluation time comparable that baseline Statistically significant word error rate (WER) reductions 0.4%-1.8% absolute (5%-11% relative) were obtained 900 h speed perturbed Switchboard corpus trained LF-MMI system multiple regularization methods including F-smoothing, L2 norm penalty, natural gradient, averaging dropout, addition i-Vector plus unit contribution (LHUC) speaker adaptation RNNLM rescoring. The efficacy is further demonstrated comparison against on same task most recent hybrid end-to-end reported literature. Consistent improvements also 450-h HKUST conversational Mandarin telephone task. On third cross domain requiring rapidly porting 1000-h LibriSpeech data small DementiaBank elderly corpus, outperformed direct fine-tuning by up 2.5% WER reduction.