Inference API error with Whisper, return_timestamps parameter

Inference Endpoints on the Hub
1

Hi team, been using the inference endpoint for whisper for months at https://api-inference.huggingface.co/models/openai/whisper-large-v3-turbo. Today, all of a sudden, the API started throwing this error

'You have passed more than 3000 mel input features (> 30 seconds) which automatically enables long-form generation which requires the model to predict timestamp tokens. Please either pass `return_timestamps=True` or make sure to pass no more than 3000 mel input features.', 'warnings': ['There was an inference error: You have passed more than 3000 mel input features (> 30 seconds) which automatically enables long-form generation which requires the model to predict timestamp tokens. Please either pass `return_timestamps=True` or make sure to pass no more than 3000 mel input features.

I tried passing the parameter in multiple ways

parameters = {
    "language": "en",
    "temperature": "0.0",
    "return_timestamps": "true",

also using 1, True (boolean), “True” But the error still comes up. I am also able to see the error on the UI playground at openai/whisper-large-v3-turbo · Hugging Face : Did anyone see similar behaviour ? Can the team please help on this ? Many thanks

Image
Image990Ă—583 89.1 KB

2 Likes
2

Hmm, I tried it with the wav file I found, and it seems to work for the moment.

whisperapi1
whisperapi1692Ă—558 28.1 KB

3

The issue is related to file length. WAV files longer than 30 seconds trigger the error.

2 Likes
4

I also faced the same problem (whisper-large-v3-turbo, ogg audio). Until yesterday it was working fine, but now audio >30s causing error, return_timestamps does not help as well. Same issue on whisper-large-v3

2 Likes
5

I agree with @TOOTHED - seeing same behaviour here. Files >30 s used to work without any issues but now throw an error, and no keyword/param works, should this be raised with the API team somehow ?

2 Likes
6

I think we can contact the Hugging Face team from any of these. The Hugging Face functions are managed on github with the exception of a few, so raising an issue on github is the most reliable way to get in touch. However, there are so many github pages that it’s hard to know where to write…:sweat_smile:

Hugging Face Documents and Hub general issue

Library handling Inference API

Whisper Class (Maybe it’s different this time…)

7

Well, tbh it is literally my first comment on hf forums, so I am not sure how it is working here. But overall, yes, as whisper models were not updated and this behaviour is similar for all them, the error is for sure caused by the updates in hugging face API - either AutomaticSpeechRecognition method (i didn’t test if any parameters work), or something deeper

1 Like
8

I have now cross-posted this in three different places - wondering if someone knows how to tag someone from the dev/support team. I feel like this is part of the broader inference api changes, the behaviour of this whisper endpoint has radically changed and is barely usable - the silence from the team on all open issues here is somewhat worrying.

1 Like
9

Hmm, I only know the staff who are visible…

If it’s related to Whisper, Serverless Inference API, or servers, then michellehbn, Wauplin, pierric, or victor?
You could also ask meganariley…
The last resort is julien-c.

10

Thanks @John6666 !

@michellehbn - tagging here as this might be related to the broader inference API issues

1 Like
11

Is there are any replacement for this model ?

1 Like
12

Maybe something like FastRTC. But the application may be a little different.