|
--- |
|
license: mit |
|
--- |
|
|
|
# Content Vec Best |
|
Official Repo: [ContentVec](https://github.com/auspicious3000/contentvec) |
|
This repo brings fairseq ContentVec model to HuggingFace Transformers. |
|
|
|
## How to use |
|
To use this model, you need to define |
|
```python |
|
class HubertModelWithFinalProj(HubertModel): |
|
def __init__(self, config): |
|
super().__init__(config) |
|
|
|
# The final projection layer is only used for backward compatibility. |
|
# Following https://github.com/auspicious3000/contentvec/issues/6 |
|
# Remove this layer is necessary to achieve the desired outcome. |
|
self.final_proj = nn.Linear(config.hidden_size, config.classifier_proj_size) |
|
``` |
|
|
|
and then load the model with |
|
```python |
|
model = HubertModelWithFinalProj.from_pretrained("lengyue233/content-vec-best") |
|
|
|
x = model(audio)["last_hidden_state"] |
|
``` |
|
|
|
## How to convert |
|
You need to download the ContentVec_legacy model from the official repo, and then run |
|
```bash |
|
python convert.py |
|
``` |
|
|
