File size: 974 Bytes
6d3f4b0 16b7417 c0b9ba1 16b7417 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 |
---
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
```
|