Nvidia Model Requests

[tbartley94]

Hi y'all,

Love the paper and leaderboard setup. Arabic dialects get short shrifted so appreciate y'alls work. Just wanted to reach out because we have some more up to date multilingual models on our end if you want more ablations comparable with Seamless and Whisper:

https://catalog.ngc.nvidia.com/orgs/nvidia/teams/riva/models/canary-riva-1b <- multilingual Transformer decoder model (Nvidia version of whisper)

https://catalog.ngc.nvidia.com/orgs/nvidia/teams/riva/models/canary-riva-0-6b-turbo <- Same but higher throughput

https://catalog.ngc.nvidia.com/orgs/nvidia/teams/riva/models/parakeet-ctc-riva-1-1b-unified-ml-cs-concat <- Multilingual variation on the CTC model you used with language specific tokenizers.

https://catalog.ngc.nvidia.com/orgs/nvidia/teams/riva/models/parakeet-ctc-riva-1-1b-unified-ml-cs-universal <- Same but tokens are merged from across all languages.

No worries about which ones you want to evaluate! Just wanted to share since it looks like we didn't release them in time for your original submission.

[BenoitWang]

Hi @tbartley94 sorry for the late reply and thanks for the info. We just acknowledge that Canary and Parakeet start to support Arabic, we are now working on it.

[tbartley94]

@BenoitWang No worries! LMK if there's anything you noticed that we're missing on our end for evaluation.

[BenoitWang]

Hi @tbartley94 Parakeets have already been added into the leaderboard, however I had an issue regarding Canary here, feel free to take a look or point out maybe something I missed, thank you.