Towards Understanding Attention-Based Speech Recognition Models

摘要

Although the attention-based speech recognition has achieved promising performances, the specific explanation of the intermediate representations remains a black box theory. In this paper, we use the method to visually show and explain continuous encoder outputs. We propose a human-intervened force alignment method to obtain labels for t-distributed stochastic neighbor embedding (t-SNE), and use them to better understand the attention mechanism and the recurrent representations. In addition, we combine t-SNE and canonical correlation analysis (CCA) to analyze the training dynamics of phones in the attention-based model. Experiments are carried on TIMIT and WSJ respectively. The aligned embeddings of the encoder outputs could form sequence manifolds of the ground truth labels. Figures of t-SNE embeddings visually show what representations the encoder shaped into and how the attention mechanism works for the speech recognition. The comparisons between different models, different layers, and different lengths of the utterance show that manifolds are clearer in the shape when outputs are from the deeper layer of the encoder, the shorter utterance, and models with better performances. We also observe that the same symbols from different utterances tend to gather at similar positions, which proves the consistency of our method. Further comparisons are taken between different epochs of the model using t-SNE and CCA. The results show that both the plosive and the nasal/flap phones converge quickly, while the long vowel phone converge slowly.