[Bug] Audio task accept headers are not respected

[StephenHodgson]

Is your feature request related to a problem? Please describe.
When settings accept headers to audio/wav or audio/mp3 when making requests to audio inference tasks using api-inference-community 0.0.32+, the returned audio file is still encoded as flac.

Describe the solution you'd like
accept headers respect the types that the client can understand.

Additional context
This is to unblock support for audio tasks in Unity game engine for upcoming hackathon.

I've tried the following tasks | models:

• text-to-speech | speechbrain/tts-tacotron2-ljspeech
• text-to-speech | espnet/kan-bayashi_ljspeech_vits
• audio-to-audio | speechbrain/sepformer-wham

I suspect that the accept headers for images likely also suffer from the same issue, but have not tested.

[Wauplin]

Hey @StephenHodgson, remind me the discussion and work done in #287, right? Looks like mp3 and wav were not in the list of accepted mime types for audio (see here). EDIT: they were added.

[StephenHodgson]

remind me the discussion and work done in #287

Sure 😊 we added support to specify a specific mime return type requested by the client for both image and audio tasks.

Looks like mp3 and wav were not in the list of accepted mime types for audio (see here).

I'm confused because that is the white listed types, and both are in the list.

I thought we had added some tests, and I assumed some tasks would lag behind the release of the api-inference-cimmunity releases. But months later testing this out with multiple different models still I'm unable to get the requested mime types back for audio tasks.