Double-attention mechanism of sequence-to-sequence deep neural networks for automatic speech recognition

Sequence-to-sequence deep neural networks with attention mechanisms have shown superior performance across various domains, where the sizes of the input and the output sequences may differ. However, if the input sequences are much longer than the output sequences, and the characteristic of the input sequence changes within a single output token, the conventional attention mechanisms are inappropriate, because only a single context vector is used for each output token. In this paper, we propose a double-attention mechanism to handle this problem by using two context vectors that cover the left and the right parts of the input focus separately. The effectiveness of the proposed method is evaluated using speech recognition experiments on the TIMIT corpus.