Effective attention-based sequence-to-sequence modelling for automatic speech recognition
With sufficient training data, attentional encoder-decoder models have given outstanding ASR results. In such models, the encoder encodes the input sequence into a sequence of hidden representations. The attention mechanism generates a soft alignment between the encoder hidden states and the decoder hidden states. The decoder produces the current output by considering the alignment and the previous outputs. However, attentional encoder-decoder models are originally designed for machine translation tasks, where the input and output sequences are relatively short and the alignments between them are flexible. For ASR tasks, the input sequences are notably long. Further, acoustic frames (or their hidden representations) typically can be aligned with output units in a left-to-right order, and compared to the length of the entire utterance, the duration of each output unit is usually small. Conventional encoder-decoder models have difficulties in modelling long sequences, and the attention mechanism does not guarantee the monotonic left-to-right alignments. In this thesis, we study attention-based sequence-to-sequence ASR models and address the aforementioned issues. We investigate recurrent neural network (RNN) encoder-decoder models and self-attention encoder-decoder models. For RNN encoder-decoder models, we develop a dynamic subsampling RNN (dsRNN) encoder to shorten the lengths of the input sequences. The dsRNN learns to skip redundant frames. Furthermore, the skip ratio may vary at different stages of training, thus allowing the encoder to learn the most relevant information for each epoch. Thus, the dsRNN alleviates the difficulties of encoding long sequences. We also propose a fully trainable windowed attention mechanism, in which both the window shift and window length are learned by the model. Our windowed method forces the attention mechanism to attend inputs within small sliding windows in a strict left-to-right order. The proposed dsRNN and windowed attention give significant performance gains over traditional encoder-decoder ASR models. We next study self-attention encoder-decoder models. For RNN encoder-decoder models, we have shown that restricting the attention within small windows is beneficial. However, self-attention encodes input sequences by comparing each element of the sequence with all other elements of the sequence. Therefore, we investigate if the global view of self-attention is necessary for ASR. We note that the range of the learned context increases from the lower to the upper self-attention layers, and suggest that the upper encoder layers may have seen sufficient contextual information without the need for self-attention. This would imply that the upper self-attention layers can be replaced with feed-forward layers (we can view the feed-forward layers as strict local left-to-right self-attention). In practice, we observe replacing upper encoder self-attention layers with feed forward layers does not impact the performance. We also observe that there are individual attention heads that only attend local information, and thus the self-attention mechanism is redundant for these attention heads. Based on these observations, we propose randomly removing attention heads during training but keep all heads at testing. The proposed method achieves state-of-the-art ASR results on benchmark datasets of different ASR scenarios. Finally, we investigate top-down level-wise training of sequence-to-sequence ASR models. We find that when training sequence-to-sequence ASR models on noisy data, the use of upper layers trained on clean data forces the lower layers to learn noise-invariant features, since the features which fit the clean-trained upper layers are more general. We further show that within the same dataset, conventional joint training makes the upper layers quickly overfit. Therefore, we propose to freeze the upper layers and retrain the lower layers. The proposed method is a general training strategy; we use it not only to train ASR models but also to train other neural networks in other domains. The proposed training method yields consistent performance gains across different tasks (e.g., language modelling, image classification). In summary, we propose methods which enable attention-based sequence-to-sequence ASR systems to better model sequential data, and demonstrate the benefits of training neural networks in a top-down cascade manner.