edit: this issue was originally about considering backends but I am hijacking it to collect my thoughts on how to (re)design everything

it would really be nice to not be in the business of maintaining a deep learning library, esp since whole teams of ppl are doing that already

would prefer scope for vak to be:

• reference implementations of algos specific to vocal learning community
• associated tools for benchmarking (e.g., WindowDataset)

question is what backends could be used that would handle training / eval / etc.
Starting this issue to keep track of thoughts

current list in my head:

• https://github.com/explosion/thinc
  • pros:
    • framework agnostic, can use torch or TF <-- for me a big pro
    • already used for spacy
    • plays well with FastAPI, pydantic, and friends
    • TOML-like configuration file format (I think, need to check back)
  • (possible) cons:
    • don't have a feeling for how easily abstractions will work outside of NLP. Need to experiment with this in a branch
• https://github.com/speechbrain/speechbrain
  • pros:
    • appear to be related abstractions we could use, e.g. loss functions
  • cons:
    • pytorch only
    • YAML config format, gross