Spaces:
Runtime error
Apply for community grant: Personal project
Hi @reach-vb @sanchit-gandhi ,
Thank for your support during this great event!
I've also fine-tuned some whisper models on German and here is the demo.
For the moment the best fine-tuned medium model got the following WERs on MCV 11.0
- 7.20% greedy
Here are the WERs reported in the paper on MCV 9.0
- 8.5% with whisper-medium
- 6.4% with whisper-large-v2
Right now the inference is taking so long on 2 CPUs, I'm wondering if you could grant GPU to this space. I will upload my large checkpoint when it's finished and wish to test it with this space.
Thanks in advance :)
Hi @sanchit-gandhi @reach-vb ,
Here are the new results of the fine-tuned models, compared with the WERs of pre-trained models reported in the paper.
Below are the WERs of the pre-trained models on the Common Voice 9.0. These results are reported in the original paper.
Model | Common Voice 9.0 |
---|---|
openai/whisper-small | 13.0 |
openai/whisper-medium | 8.5 |
openai/whisper-large-v2 | 6.4 |
Below are the WERs of the fine-tuned models on the Common Voice 11.0.
Model | Common Voice 11.0 |
---|---|
bofenghuang/whisper-small-cv11-german-punct | 11.35 |
bofenghuang/whisper-medium-cv11-german-punct | 7.05 |
bofenghuang/whisper-large-v2-cv11-german-punct | 5.76 |
@bofenghuang
- WOW! This is brilliant. Lovely to see this work so well :)
We've assigned you a GPU. The grant will go on till the end of the week. All the best!