Transformer-based Mannequin for ASR N-Greatest Rescoring and Rewriting

Transformer-based Mannequin for ASR N-Greatest Rescoring and Rewriting
Transformer-based Mannequin for ASR N-Greatest Rescoring and Rewriting


Voice assistants more and more use on-device Automated Speech Recognition (ASR) to make sure pace and privateness. Nevertheless, as a consequence of useful resource constraints on the gadget, queries pertaining to advanced data domains usually require additional processing by a search engine. For such functions, we suggest a novel Transformer based mostly mannequin able to rescoring and rewriting, by exploring full context of the N-best hypotheses in parallel. We additionally suggest a brand new discriminative sequence coaching goal that may work effectively for each rescore and rewrite duties. We present that our Rescore+Rewrite mannequin outperforms the Rescore-only baseline, and achieves as much as a median 8.6% relative Phrase Error Charge (WER) discount over the ASR system by itself.

Leave a Reply

Your email address will not be published. Required fields are marked *