How to pass arguments to model when using InferenceClient

Inference Endpoints on the Hub
1

I would like to use Meta’s MMS model with the InferenceClient. However it supports multiple languages and I need to pass the following keywords:

{"target_lang":"spa", "ignore_mismatched_sizes":True}

How can I do this? Here is my code:

from huggingface_hub import InferenceClient
client = InferenceClient()
client.automatic_speech_recognition("test.wav", model="facebook/mms-1b-all").text

It’s returning a result, but it clearly doesn’t recognize which language is being spoken.

2

cc @Wauplin

3

Hi @khof312 , this is unfortunately not officially supported by InferenceClient. You can do something like this:

from huggingface_hub import AutomaticSpeechRecognitionOutput, InferenceClient
from huggingface_hub.inference._common import _b64_encode


client = InferenceClient()
response = client.post(
    json={
        "inputs": _b64_encode("test.wav"),
        "parameters": {
            "target_lang": "spa",
            "ignore_mismatched_sizes": True,
        },
    },
    model="facebook/mms-1b-all",
)
output = AutomaticSpeechRecognitionOutput.parse_obj_as_instance(response)

but remember this is not an official support.

To add new inputs and parameters to the InferenceClient, you first need to add them to the “official” tasks specification here. Only popular parameter are added to avoid bloating the API (and constraining ourselves for future breaking changes).