wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations
Paper • 2006.11477 • Published • 10
# Load model directly
from transformers import AutoProcessor, AutoModelForCTC
processor = AutoProcessor.from_pretrained("OthmaneJ/distil-wav2vec2")
model = AutoModelForCTC.from_pretrained("OthmaneJ/distil-wav2vec2", device_map="auto")Quick Links
Distil-wav2vec2
This model is a distilled version of the wav2vec2 model (https://arxiv.org/pdf/2006.11477.pdf). This model is 45% times smaller and twice as fast as the original wav2vec2 base model.
Evaluation results
This model achieves the following results (speed is mesured for a batch size of 64):
| Model | Size | WER Librispeech-test-clean | WER Librispeech-test-other | Speed on cpu | speed on gpu |
|---|---|---|---|---|---|
| Distil-wav2vec2 | 197.9 Mb | 0.0983 | 0.2266 | 0.4006s | 0.0046s |
| wav2vec2-base | 360 Mb | 0.0389 | 0.1047 | 0.4919s | 0.0082s |
Usage
notebook (executes seamlessly on google colab) at https://github.com/OthmaneJ/distil-wav2vec2
- Downloads last month
- 71
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("automatic-speech-recognition", model="OthmaneJ/distil-wav2vec2")