|
|
--- |
|
|
license: mit |
|
|
datasets: |
|
|
- google/speech_commands |
|
|
base_model: |
|
|
- facebook/wav2vec2-base |
|
|
--- |
|
|
# Command Classifier |
|
|
|
|
|
This is a lightweight classifier model trained on top of Wav2Vec2 features for classifying speech commands. |
|
|
|
|
|
- **Base model**: `facebook/wav2vec2-base` |
|
|
- **Fine-tuned model**: Lightweight linear classifier |
|
|
- **Dataset**: [Google Speech Commands v0.02](https://arxiv.org/abs/1804.03209) |
|
|
|
|
|
## Usage |
|
|
|
|
|
You can use this classifier by combining it with Wav2Vec2 features. The classifier expects mean-pooled Wav2Vec2 hidden states. |
|
|
|
|
|
```python |
|
|
from transformers import Wav2Vec2Model |
|
|
from command_classifier import CommandClassifier |
|
|
import torch |
|
|
|
|
|
wav2vec = Wav2Vec2Model.from_pretrained("facebook/wav2vec2-base") |
|
|
classifier = CommandClassifier(num_classes=35) |
|
|
classifier.load_state_dict(torch.load("pytorch_model.bin")) |
|
|
|
