Lightweight domain adaptation: A filtering pipeline to improve accuracy of an Automatic Speech Recognition (ASR) engine

Transformer models have accelerated the field of speech recognition; deriving a low word error rate (WER) is demonstrably achievable under varying conditions. However, most ASR engines are trained on acoustic and language models constructed from corpora that include news feeds, books, and blogs in order to demonstrate generalization, leading to errors when the model is applied to a specific domain. While the increase in WER is acute for very specific domains (health and medicine), our work shows that it is sizable even when the domain is general (hospitality). For such domains, a lightweight adaptation approach can help; lightweight because the adaptation does not require extensive post-hoc training of additional domain-specific acoustic- or language-models that act as adjutants to the base ASR engine. We present our work on such lightweight filtering pipeline that seamlessly integrates lightweight models (n − gram, decision trees) with powerful, pre-trained, bi-directional transformer models, all working in conjunction to derive a 1-best hypothesis word selection algorithm. Our pipeline reduces the WER between 1.6% to 2.5% absolute while treating the ASR engine as a black box, and without requiring additional complex discriminative training.