Imputer: Sequence Modelling via Imputation and Dynamic Programming
This addresses the computational bottleneck in sequence-to-sequence tasks like speech recognition, offering a faster alternative to autoregressive models with competitive accuracy.
The paper tackles the problem of slow sequence generation in neural models by introducing the Imputer, which iteratively imputes output sequences with a constant number of steps, achieving a word error rate of 11.1 on LibriSpeech test-other, outperforming prior non-autoregressive models.
This paper presents the Imputer, a neural sequence model that generates output sequences iteratively via imputations. The Imputer is an iterative generative model, requiring only a constant number of generation steps independent of the number of input or output tokens. The Imputer can be trained to approximately marginalize over all possible alignments between the input and output sequences, and all possible generation orders. We present a tractable dynamic programming training algorithm, which yields a lower bound on the log marginal likelihood. When applied to end-to-end speech recognition, the Imputer outperforms prior non-autoregressive models and achieves competitive results to autoregressive models. On LibriSpeech test-other, the Imputer achieves 11.1 WER, outperforming CTC at 13.0 WER and seq2seq at 12.5 WER.