Audio Data Loader Example?

Hi, I'd like to use FairSeq2 for my audio models. Is there an example of an audio data loader in the works? That would be super helpful.

What I need in the dataloader:
1. I have lots of audio that I'd like to load. This is currently in mp3, but I've been cutting chunks out as tuples of (audio: Wav, labels: Json).
2. This data is hosted on R2 and I'd like the loader to stream it via an endpoint_url.
3. I also would like for this to save to local cache as it's a big dataset.
4. Distribute amongst cpu workers.

I can accommodate other data saving formats as needed, e.g. a flat zip, but the most important thing is that it's fast. I'm right now I/O bottlenecked.

Thanks!