Anyone who has shipped an ASR product knows the pain of proper nouns: contact names, song titles, medication names, obscure place names. CLAS, or Contextual Listen-Attend-Spell, is Google's answer for injecting that context directly into an end-to-end speech recognizer. Instead of biasing beam search after the fact, CLAS learns an attention mechanism over a list of bias phrases, so the model can lean on user-specific vocabulary at decode time without retraining.

This long review breaks down the architecture, the training tricks that make bias attention robust to distractors, and the WER wins on realistic voice-assistant tasks. If you're building on-device ASR, dictation apps, or voice interfaces where personalization matters, CLAS remains a foundational read even years later, and its ideas keep showing up in modern RNN-T and CTC contextual biasing work. Pull up a notebook and follow along if contextual biasing is on your ASR roadmap.