huggingface_hub

70e66119 - Upgrading community frameworks to `audio-to-audio`. (#94)

Upgrading community frameworks to `audio-to-audio`. (#94) * Add `audio-to-audio` and remove `audio-source-separation` to this package. * Adding `audio-to-audio` task. - Removed `audio-source-separation` (it's not `audio-to-audio` and it is a bit more general). Also the previous version was flawed as there was no way to return multiple channels. - Decision to return base64 blobs in JSON. It's not super shiny but it does the work and enabled easy ways to add annotations for audio streams (like the instrument labels for source separation). - Added GZipMiddleware within common, this help base64 encoded json be omewhat ok in terms of size. - Known caveat. Currently we only send 1 channel to the input of the pipeline. At the moment of this commit, no model requires stereo or multi channel input, so keeping that change for later. - community framework update will follow. - Because of dockerized versionning we can safely delete `audio-source-separation`. Old `api-inference-community` will continue to support those and the hub too. But if you want to use > 0.0.7 then you will need to migrate. * Adding type checks in tests for `audio-to-audio`. * Upgrading community frameworks to `audio-to-audio`. - Deleted `audio-source-separation` as much as possible. Necessary to keep some legacy code, because the `audio-source-separation` tag and widget still exist. - Targeted `speechbrain` and `asteroid` the only 2 frameworks to implement `audio-source-separation` at this point. Reference: https://github.com/huggingface/huggingface_hub/pull/76 * Rebase. Co-authored-by: cem <csubakan@gmail.com> * Fixing black. * isort. * Upgrading espnet + speechbrain to fallable start. Co-authored-by: cem <csubakan@gmail.com>