Reproduce “Deep context: end-to-end contextual speech recognition” on OpenResearch

About this paper

Golan Pundak, Tara N. Sainath, Rohit Prabhavalkar, Anjuli Kannan, Ding Zhao

Google Inc. researchers developed DEEP CONTEXT, an end-to-end contextual speech recognition system (CLAS) that natively incorporates dynamic context phrases into an attention-based neural network. This architecture demonstrated up to a 68% relative Word Error Rate reduction on tasks involving hundreds of context phrases and successfully scaled to thousands of phrases with a 'bias-conditioning' mechanism.

Key insights

• Jointly optimizing contextual information directly within an end-to-end ASR model through an attention mechanism can lead to substantial accuracy improvements over traditional shallow fusion methods, especially for out-of-vocabulary terms.
• The proposed Contextual-LAS (CLAS) architecture, even without explicit context, demonstrates increased robustness compared to a vanilla LAS model, suggesting benefits from its training regimen.
• Scalability is a critical challenge for contextual ASR with very large context lists (thousands of phrases), but a 'Bias-Conditioning' mechanism, which activates bias phrases based on partial hypothesis matches, effectively mitigates this issue.

Summary by alphaXiv.