add support for Kokoro via Replicate

[Vaibhavs10]

No description provided.

[julien-c]

need to add a test but you'll see, you'll need to change the param name

I was meaning to ask @jbilcke-hf to update the param name in jbilcke/oute-tts to text instead of inputs

[hanouticelina]

Not sure if this will work, there is an input payload schema mismatch between text-to-speech models in Replicate.
While jbilcke/oute-tts expects the prompt in input["inputs"], jaaari/kokoro-82m requires it in input["text"] and the implementation currently uses input["inputs"] (see here).

[Vaibhavs10]

it was too easy to be true, yeah will update the cog ID to a different to match the spec!

[julien-c]

yes, imo you can remove the oute-tts one for now (but keep the test as skipped and ask if @jbilcke-hf can update his replicate model in the future) and change the param name to text

[julien-c]

https://replicate.com/jbilcke/oute-tts was not Warm anyways so not great UX (a bit slow to start)

[hanouticelina]

yes, imo you can remove the oute-tts one for now (but keep the test as skipped and ask if @jbilcke-hf can update his replicate model in the future) and change the param name to text

I'll take care of doing the equivalent Python side. Also, it seems to me that we have more text-to-speech models using the text parameter instead of inputs, eg: we will be able to add xtts-v2 (replicate link, HF link) as well.
EDIT: xtts-v2 does not support the same API either, since it's a text-to-speech voice cloning model.

[julien-c]

Approving, modulo my comment

[SBrandeis]

Does it accept parameters?
Would be nice to add some to make sure they payload has the proper shape

[hanouticelina]

yes, but it does not accept any of the parameters listed in TextToSpeechParameters (see the schema of the model here). Since each model on Replicate runs in its own docker container, models of the same task can have different parameters (and even more problematic: different input/output format). In huggingface_hub, we added an argument extra_body to allow users to pass any provider- or model-specific parameters.

[julien-c]

i'll merge this now, to be able to test it on the Hub 🔥

[julien-c]

oops, i did not add the VCR tape => #1171