Microsoft WavLM-Base-Plus for Speaker Verification is corrupted

Models
1

Hi there,

I’ve just tried the demo code from this page: microsoft/wavlm-base-plus-sv · Hugging Face on Colab.

However, it’s giving this corruption warning:

/usr/local/lib/python3.7/dist-packages/transformers/modeling_utils.py in _load_state_dict_into_model(cls, model, state_dict, pretrained_model_name_or_path, ignore_mismatched_sizes, _fast_init)
   1680             if any(key in expected_keys_not_prefixed for key in loaded_keys):
   1681                 raise ValueError(
-> 1682                     "The state dictionary of the model you are training to load is corrupted. Are you sure it was "
   1683                     "properly saved?"
   1684                 )

ValueError: The state dictionary of the model you are training to load is corrupted. Are you sure it was properly saved?

I also tried the live demo from this page: WavLM Speaker Verification - a Hugging Face Space by microsoft , but when I tried recording my speech and submit, it returned an unknown error.

Is there anyone that could help get arround this?

Thanks!

2

Thanks for reporting this! @anton-l could you take a look here?

1 Like
3

Hey @lamnguyenx, thank you for reporting those errors!

The first one is due to an outdated example for wavlm-base-plus-sv: it should use WavLMForXVector to load the model. The example is now fixed :hugs: microsoft/wavlm-base-plus-sv · Hugging Face

The error in the live demo is related to the recent changes in the Gradio Audio input. We’ve fixed the error in that particular demo and are working on an upstream fix :slight_smile:

1 Like
4

Thank you, @anton-l !
It’s working perfectly now :grin: