Low-rank Adaptation of Large Language Model Rescoring for Parameter-Efficient Speech Recognition
This work addresses efficiency and domain adaptation problems for speech recognition practitioners, representing an incremental improvement over existing rescoring methods.
The paper tackled the high computational cost and domain adaptation limitations of pretrained language models like BERT in speech recognition rescoring by proposing a low-rank adaptation method, resulting in training time reductions by factors of 5.4 to 3.6 while using only 0.08% of pretrained parameters.
We propose a neural language modeling system based on low-rank adaptation (LoRA) for speech recognition output rescoring. Although pretrained language models (LMs) like BERT have shown superior performance in second-pass rescoring, the high computational cost of scaling up the pretraining stage and adapting the pretrained models to specific domains limit their practical use in rescoring. Here we present a method based on low-rank decomposition to train a rescoring BERT model and adapt it to new domains using only a fraction (0.08%) of the pretrained parameters. These inserted matrices are optimized through a discriminative training objective along with a correlation-based regularization loss. The proposed low-rank adaptation Rescore-BERT (LoRB) architecture is evaluated on LibriSpeech and internal datasets with decreased training times by factors between 5.4 and 3.6.