-
Notifications
You must be signed in to change notification settings - Fork 2.1k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Libri100 recipe for standalone Transducer #4698
base: master
Are you sure you want to change the base?
Conversation
Codecov Report
@@ Coverage Diff @@
## master #4698 +/- ##
==========================================
+ Coverage 74.36% 74.99% +0.63%
==========================================
Files 654 655 +1
Lines 58347 58546 +199
==========================================
+ Hits 43391 43909 +518
+ Misses 14956 14637 -319
Flags with carried forward coverage won't be shown. Click here to find out more. see 15 files with indirect coverage changes 📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more |
--ngpu 1 \ | ||
--nj 32 \ | ||
--inference_nj 32 \ | ||
--nbpe 500 \ |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
BPE size is 500
vs 5000
for the CTC-Att baseline model. I guess we can further improve results with bigger BPE size but I don't have enough resource for that.
This pull request is now in conflict :( |
It's more stable on my side so I'll start training different model architectures and settings before opening the corresponding PR. I'll use this branch (and message) to keep track of the best models.
|
for more information, see https://pre-commit.ci
@sw005320 If we are okay with the design (i.e.: adding a Pre-trained model links will be added after I finish experiments with RWKV and I need to re-train the models anyways. At least, users have some examples for the standalone version and Librispeech-100 for now (which was asked before). |
Add scripts + configs for streaming and offline Transducer. The second model is training, I'll add results to the README and update both models to hf.
Also, I added some minor fixes to the PR, it won't impact previous versions in any way.I made a separate directory for
asr_transducer1
here because we need some separation between the two versions of Transducer in ESPnet2 but I don't really have an opinion on the design. It makes sense because we defined this version as a new "task" but we could also keepasr1
and use some prefixes for the standalone version.