Locality enhanced dynamic biasing and sampling strategies for contextual ASR
This work addresses incremental improvements in contextual biasing for ASR, specifically for handling rare phrases in speech recognition systems.
The paper tackled the challenge of recognizing time-variant rare phrases in Automatic Speech Recognition by analyzing sampling strategies for contextual biasing and introducing a neighbourhood attention mechanism to refine outputs, resulting in an average 25.84% relative WER improvement on LibriSpeech sets and rare-word evaluation compared to the baseline.
Automatic Speech Recognition (ASR) still face challenges when recognizing time-variant rare-phrases. Contextual biasing (CB) modules bias ASR model towards such contextually-relevant phrases. During training, a list of biasing phrases are selected from a large pool of phrases following a sampling strategy. In this work we firstly analyse different sampling strategies to provide insights into the training of CB for ASR with correlation plots between the bias embeddings among various training stages. Secondly, we introduce a neighbourhood attention (NA) that localizes self attention (SA) to the nearest neighbouring frames to further refine the CB output. The results show that this proposed approach provides on average a 25.84% relative WER improvement on LibriSpeech sets and rare-word evaluation compared to the baseline.