Improving Rare Word Recognition with LM-aware MWER Training
This work addresses the training-inference gap in language model usage for speech recognition, offering incremental improvements for systems handling rare words.
The paper tackled the problem of rare word recognition in end-to-end models by integrating language models into the discriminative training of hybrid autoregressive transducer models, achieving a 10% relative improvement on voice search test sets with rare words and eliminating the need for manual fusion weight tuning in rescoring setups.
Language models (LMs) significantly improve the recognition accuracy of end-to-end (E2E) models on words rarely seen during training, when used in either the shallow fusion or the rescoring setups. In this work, we introduce LMs in the learning of hybrid autoregressive transducer (HAT) models in the discriminative training framework, to mitigate the training versus inference gap regarding the use of LMs. For the shallow fusion setup, we use LMs during both hypotheses generation and loss computation, and the LM-aware MWER-trained model achieves 10\% relative improvement over the model trained with standard MWER on voice search test sets containing rare words. For the rescoring setup, we learn a small neural module to generate per-token fusion weights in a data-dependent manner. This model achieves the same rescoring WER as regular MWER-trained model, but without the need for sweeping fusion weights.