further refinement to the swbd recipe

[JinZr]

including a narrowband setup

[huangruizhe]

Thanks for the efforts! Do you know what's the WER of the new recipe, and will there be a trained model released? Thank you.

[JinZr]

hi!

I have updated the WERs of narrowband systems in the latest commit. This PR should be merged soon, thanks for the waiting!

On Wed, Nov 29, 2023 at 11:54 PM huangruizhe @.***> wrote:

Thanks for the efforts! Do you know what's the WER of the new recipe, and will there be a trained model released? Thank you.

— Reply to this email directly, view it on GitHub https://github.com/k2-fsa/icefall/pull/1391#issuecomment-1832183366, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOON42EKC4GT3YQE2UZQM2TYG5LCLAVCNFSM6AAAAAA7YRWL5SVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQMZSGE4DGMZWGY . You are receiving this because you authored the thread.Message ID: @.***>

[huangruizhe]

Thanks. Is it true that the WER degrades quite a lot after the narrow band setup? Have you used MUSAN augmentation in the latest setup?

[JinZr]

Yes, indeed. I am a bit confused about the results as well, I think we can certainly arrange a similar experiment using ESPnet with a few tweaks to the setup (they upsample the data to 16kHz as a default) and see how the results change 🧐.