Using External Off-Policy Speech-To-Text Mappings in Contextual End-To-End Automated Speech Recognition