add GigaSpeech dataset in SpeechBrain #2405
Draft
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Work in progress. Do not review/merge.
This PR aims at adding GigaSpeech dataset inside of SpeechBrain.
Webdataset support ?
At first, I designed the data prep script so that it supports out of the box webdataset. However, I found in my experiments that the way we were using webdataset was not optimal. Indeed, and as described in the Google Colab tutorial, we first create shards and then apply our traditional padding function, etc. The issue with that is that the shards may contain audio files with very long sequences and very short sequences resulting in a lot of padding. Furthermore, many of our SpeechBrain features are not working easily with shards. For instance, training a simple label encoder will require some engineering tricks to make it work on shards. Padding as well. Also, webdataset has evolved and it would require us to freeze the dependence on a very old version of webdataset. I had some time to look at other toolkit implementations and found that NeMo, for instance, was first sorting audio files, and THEN creating shards. I also found that Lhoste has its own webdataset implementation which seems more tailored to the speech modality. (Maybe we should have a closer look). Thus, I decided temporarily to remove webdataset from this PR. I think most of the people that will be using GigaSpeech XL split will have access to more than 1 TB of storage and it won't be an issue at all. I am open to discussion but it will require some design discussion.
General Todo
To do:
parallel_map
so that it is very fastCTC
To do:
Maybe one day:
Reference
k2 Icefall
Model | Dev | Test
zipformer | 10.25 | 10.38
conformer_ctc | 10.47 | 10.58
pruned_transducer_stateless2 | 10.40 | 10.51
Before submitting
PR review
Reviewer checklist