LT-LM: a novel non-autoregressive language model for single-shot lattice rescoring

Neural network-based language models are commonly used in rescoring\napproaches to improve the quality of modern automatic speech recognition (ASR)\nsystems. Most of the existing methods are computationally expensive since they\nuse autoregressive language models. We propose a novel rescoring approach,\nwhich processes the entire lattice in a single call to the model. The key\nfeature of our rescoring policy is a novel non-autoregressive Lattice\nTransformer Language Model (LT-LM). This model takes the whole lattice as an\ninput and predicts a new language score for each arc. Additionally, we propose\nthe artificial lattices generation approach to incorporate a large amount of\ntext data in the LT-LM training process. Our single-shot rescoring performs\norders of magnitude faster than other rescoring methods in our experiments. It\nis more than 300 times faster than pruned RNNLM lattice rescoring and N-best\nrescoring while slightly inferior in terms of WER.\n