Skip to content

New ASR dataset VoxPopuli and Transducer recipe (Code from Samsung AI Center Cambridge) - #2421

Merged
Adel-Moumen merged 57 commits into
speechbrain:developfrom
TParcollet:voxpopuli_transducer
Mar 21, 2024
Merged

Adel-Moumen merged 57 commits into
speechbrain:developfrom
TParcollet:voxpopuli_transducer

Conversation

@TParcollet

@TParcollet TParcollet commented Feb 20, 2024

Copy link
Copy Markdown
Collaborator

What does this PR do?

This PR introduces a streaming ASR recipe for VoxPopuli with a transducer. I'll try to commit a transformer and CTC (SSL) recipe as well ... later.

To-do

  • add tests
  • A reviewer must retrain the model on 2 - 4 GPUs (batch size between 800s and 1000s)

Titouan Parcollet/Embedded AI /SRUK/Engineer/Samsung Electronics added 2 commits February 20, 2024 14:17
@TParcollet TParcollet added enhancement New feature or request ready to review Waiting on reviewer to provide feedback labels Feb 21, 2024
@TParcollet

Copy link
Copy Markdown
Collaborator Author

@Adel-Moumen and @asumagic ready to review

@mravanelli

Copy link
Copy Markdown
Collaborator

@Adel-Moumen, do you think you can proceed with the review? It would be great to include it in the next release if everything is fine.

@Adel-Moumen

Copy link
Copy Markdown
Collaborator

@Adel-Moumen, do you think you can proceed with the review? It would be great to include it in the next release if everything is fine.

I'll do the review, but note that It won't come in the next release. I won't have enough time to retrain everything/host on HF/make a good review in a short amount of time. SB 1.0 is already huge and we should now think about the future, and this with GigaSpeech should be merged in the upcoming weeks so that we still have news to share with the community.

@Adel-Moumen Adel-Moumen left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Dear @TParcollet,

Thanks for your contribution. I left a review, most of my comments concerns some little improvements such as docstring/data prep.

I will now start to train the model. I plan to update the README asap and host an HF streaming model as well.

BTW, please fix the tests ^-^

Comment thread recipes/VoxPopuli/ASR/transducer/train.py Outdated
Comment thread recipes/VoxPopuli/ASR/transducer/hparams/conformer_transducer.yaml
Comment thread recipes/VoxPopuli/ASR/transducer/train.py Outdated
Comment on lines +3 to +4
Download: https://github.com/facebookresearch/voxpopuli
Author

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

a return line would be great :p

Comment thread recipes/VoxPopuli/ASR/voxpopuli_prepare.py
Comment thread recipes/VoxPopuli/ASR/voxpopuli_prepare.py
Comment thread recipes/VoxPopuli/ASR/transducer/train.py
Comment thread recipes/VoxPopuli/Tokenizer/train.py
Comment thread recipes/VoxPopuli/Tokenizer/voxpopuli_prepare.py Outdated
Comment thread tests/recipes/VoxPopuli.csv
@TParcollet

TParcollet commented Mar 18, 2024

Copy link
Copy Markdown
Collaborator Author

Tests are failing with an obscure error that seems to be unrelated to this PR? @Adel-Moumen

@Adel-Moumen Adel-Moumen left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

lgtm.

@Adel-Moumen
Adel-Moumen merged commit 35ef78c into speechbrain:develop Mar 21, 2024
pchampio pushed a commit to pchampio/speechbrain that referenced this pull request Apr 17, 2024
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

enhancement New feature or request ready to review Waiting on reviewer to provide feedback

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants