I wanted to get people’s thoughts on merging the fastai audio library here into fastai. I have done this locally and there is a lot of work to do still in developing, but it’s a really good start to a fastai.audio implementation. I believe this would help the audio library efforts because it would bring everybody’s work back into the main fastai library.
3 Likes
Hey Kevin, this is what we’re working towards and hope it will eventually be merged in to the main library, but at the moment it’s not totally stable. We have a few known issues we need to resolve but have been waiting for fastai v2 to release since it will have breaking changes. At that point we aim to do a refactor and fix those problems that we haven’t wanted to fix yet because v2 will likely cause us to need another refactor. We also have very limited testing (no sense in writing tests that will need to be rewritten very soon).
We would love to have some new devs helping us to get it into shape for merging, especially those who are more experienced with fastai, but all skill levels are welcome. Does anyone have a ballpark date for when v2 will be released?
1 Like
That’s the best timeline that I can find for v2. I have to think it’s super close to release but I would have guessed that beginning of July as well. I guess we might just need to continue to be patient on this and work on getting audio either added in v3 or v2.1. I am planning on continuing to work on getting the audio library ready for prime-time. Basically just keep using it and as I find bugs or things that aren’t implemented that would be nice for me, I will develop them and create pull requests.
I’ve been really impressed with the work that is already done on it. The area I’m mostly focused on at the moment is Automatic Speech Recognition using the same strategy that DeepSpeech uses inside of fastai.
1 Like
Thanks, it means a lot that you’ve found it useful. My contribution process has been similar to what you mentioned. I work on a project and as I come across ideas/features that I think would be generally useful, I switch over to doing library code and implement them so that other people don’t have to reinvent the wheel. It’s been a great learning experience and I’m looking forward to hearing your ideas.