English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Discriminative training techniques define state-of-the-art performance forautomatic speech recognition systems. However, they are inherently prone tooverfitting, leading to poor generalization performance when using limitedtraining data. In order to address this issue, this paper presents a fullBayesian framework to account for model uncertainty in sequence discriminativetraining of factored TDNN acoustic models. Several Bayesian learning based TDNNvariant systems are proposed to model the uncertainty over weight parametersand choices of hidden activation functions, or the hidden layer outputs.Efficient variational inference approaches using a few as one single parametersample ensure their computational cost in both training and evaluation timecomparable to that of the baseline TDNN systems. Statistically significant worderror rate (WER) reductions of 0.4%-1.8% absolute (5%-11% relative) wereobtained over a state-of-the-art 900 hour speed perturbed Switchboard corpustrained baseline LF-MMI factored TDNN system using multiple regularizationmethods including F-smoothing, L2 norm penalty, natural gradient, modelaveraging and dropout, in addition to i-Vector plus learning hidden unitcontribution (LHUC) based speaker adaptation and RNNLM rescoring. Consistentperformance improvements were also obtained on a 450 hour HKUST conversationalMandarin telephone speech recognition task. On a third cross domain adaptationtask requiring rapidly porting a 1000 hour LibriSpeech data trained system to asmall DementiaBank elderly speech corpus, the proposed Bayesian TDNN LF-MMIsystems outperformed the baseline system using direct weight fine-tuning by upto 2.5\% absolute WER reduction.

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