eduzhai > Applied Sciences > Engineering >

Black-box Adaptation of ASR for Accented Speech

  • Save

... pages left unread,continue reading

Document pages: 5 pages

Abstract: We introduce the problem of adapting a black-box, cloud-based ASR system tospeech from a target accent. While leading online ASR services obtainimpressive performance on main-stream accents, they perform poorly onsub-populations - we observed that the word error rate (WER) achieved byGoogle s ASR API on Indian accents is almost twice the WER on US accents.Existing adaptation methods either require access to model parameters oroverlay an error-correcting module on output transcripts. We highlight the needfor correlating outputs with the original speech to fix accent errors.Accordingly, we propose a novel coupling of an open-source accent-tuned localmodel with the black-box service where the output from the service guidesframe-level inference in the local model. Our fine-grained merging algorithm isbetter at fixing accent errors than existing word-level combination strategies.Experiments on Indian and Australian accents with three leading ASR models asservice, show that we achieve as much as 28 relative reduction in WER overboth the local and service models.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×