TL;DR: Use torchaudio, it is a well-maintained, mostly stable library written by extremely skilled (and active) engineers.
Hey sorry to all here who have become interested in FastAudio and see that it is no longer maintained. I know how frustrating it is to see a potentially useful open source project and then notice the last commits are more than a year out and recent issues are unresolved.
I can confirm @KevinB’s reasoning here. In addition to the reasons he listed, torchaudio has also evolved massively since the days of fastaudio, replacing many of the reasons we started the project in the first place. For instance, when we began, torchaudio didn’t provide feature extractions like spectrograms/mfccs, and getting those things to work on a GPU was sorely needed. Finally, because these things weren’t abstracted away, it was important to learn the details of signals processing, and treat things like n_mel and the number of bins as hyperparams, something that is much less necessary in 2024.
If I were a new person getting into audio today I would
- Look for a pretrained model that covers my use case out of the box
- If that doesn’t work, look for a pretrained model that I could finetune on my data using torchaudio and the deep learning framework of your choice
- Try to not get bogged down learning the technical details of audio/signals processing unless it interests you, or you can find no other way forward in your work/research. You can get quite far without it.