microsoft

/

wavlm-large

like 95

Follow

Microsoft 18.1k

I was examining the FINALLY speech enhancement model (https://arxiv.org/abs/2410.05920) released by SamsungLabs, which uses WavLM-Large as part of its architecture. I noticed a discrepancy in the reported number of parameters:

• The original WavLM-Large model on Hugging Face is documented to have around 316M parameters.
• However, the FINALLY model description mentions a larger number of parameters (358M parameters) attributed to the WavLM component, which does not match the standard WavLM-Large count.
  

I’m curious if this is due to:

• A modified WavLM-Large backbone in FINALLY (e.g., some layers added),
• A different counting method (trainable vs total parameters), or
• A documentation oversight.

Has anyone else noticed this mismatch? It would be helpful to clarify for anyone trying to reproduce results or analyze computational requirements.

Thanks in advance!