Inquiry about WavLM German Model Variants and Their Performance
Hello,
I've noticed that three variants of the WavLM model for the German language have been released without accompanying evaluation results. I took the initiative to evaluate all three variants on the MCV 7.0 Test Dataset, and here are the results:
Variant WER CER
s101 50.61% 12.85%
s295 50.03% 12.68%
s824 48.86% 12.31%
Given that these results lag behind other models, such as Wav2Vec 2.0, I was hoping you could provide more information on how these WavLM models were trained and what the key differences are between them. The literature suggests that WavLM has an edge over Wav2Vec 2.0, at least for the English variants, so I’m curious about why this advantage doesn’t seem to translate to the German versions.
Thank you for any insights you can provide.