Bottleneck linear transformation network adaptation for speaker adaptive training-based hybrid DNN-HMM speech recognizer

Automatic node selection for Deep Neural Networks using Group Lasso regularization

“…However, a large network is obviously undesirable from the viewpoints of computational load and memory size; it is also unfavorable from the viewpoint of controlling training robustness to unseen data. To meet this requirement for finding a small, necessary, and sufficient DNN structure, several approaches have reshaped the network structure [3,4,5] or pruned the network nodes [6]. However, these methods assumed retraining or adapting a size-reduced network for high discriminative power.…”

Speaker Adaptive Training Using Model Agnostic Meta-Learning

“…In model-based speaker adaptive training this may be done by splitting the weights of the acoustic model into a speaker-independent and a speaker-dependent set. During training, a copy of the speaker-dependent weights is maintained and optimised for each speaker separately [12,13,14]. Here, we take an alternative approach: Instead of maintaining and optimising a separate copy of speaker-dependent weights for each speaker we embed speaker adaptation directly into the acoustic model training using a meta-learning approach in order to find a good initialisation for speaker-dependent weights.…”

“…In model-based speaker adaptive training, the acoustic model is parameterised as speakerdependent and speaker-independent weights. A copy of the speaker-dependent weights is maintained and optimised separately for each speaker during the training process in order to factor out speaker variation from the canonical speakerindependent acoustic model [12,13,14]. Finally, all hybrid approaches can be considered as speaker adaptive training because they provide information about speaker identity, which allows the acoustic model to easily remove speaker variation from the input features [9,10,15].…”