Faster loading of sounds

[JorisCos]

Hey Iver,

I’ve been using audiomentations for dynamic mixing in a DNN pipeline for speech enhancement and source separation, and I ran into some loading-related bottlenecks.

I made a few changes on my side that speed things up quite a bit by only loading the audio that's actually needed. The trade-off is that lru_cache doesn't really work with this approach anymore, since the access pattern becomes less predictable.

Would you be open to a PR or discussion around this?

Cheers

[iver56]

Hey Joris! Long time no speak. Is it AddBackgroundNoises you are talking about here? There's an open PR that proposes to speed up loading by loading only the part that is needed: #360
It's somewhere on my TODO list, it just hasn't gotten to the top of my list of priorities yet 😅

Since it complicates lru_cache, I would prefer to remove the lru_cache altogether, because most of the time the noise dataset is larger than what is sensible to store in RAM anyway, and the loading won't become a bottleneck as long as it gets loaded from a sufficiently fast NVMe.

[iver56]

@JorisCos How does that PR (#360) compare to what you have done in your private fork? Want to help review the PR? Or share your solution, for reference?

[JorisCos]

I actually missed this PR thanks for pointing it out ! I came up with the same answer as the proposed one. I also extended it to ApplyImpulseResponse although the performance impact is less significant with this augmentation.
I would be glad to review and share I think we can close this issue and I will contribute to #360.

[iver56]

Thanks!

[iver56]

@JorisCos The PR is merged and the LRU cache has been removed. Does it look alright? Could you install the bleeding edge version (the main branch) of audiomentations on your end and do a training run to test it?

pip install git+https://github.com/iver56/audiomentations.git@main

[JorisCos]

Yes it looks good, unfortunately I'm in between contracts and I currently don't have access to computing resources nor datasets so I can't test in a real scenario sry.

[worldveil]

In cases where memory can be traded for processing time, this PR creates a shared cache for more quickly loading audio: #419