Wav2vec2 in PaddlePaddle
This is paddle-paddle version of Facebook's Wav2vec2.0 [1], with code and pre-trained weighted ported from Fairseq and huggingface.
Install PaddlePaddle 2.0.1
pip install PaddlePaddle-gpu==2.0.1
Install PaddleAudio by
git clone https://github.com/PaddlePaddle/models
cd models/PaddleAudio
pip install -e .
| name | Finetuning split | Dataset |
|---|---|---|
| wav2vec2-base-960h | 960h | Librispeech |
| wav2vec2-large-960h | 960h | Librispeech |
| wav2vec2-base-960h-lv60 | 960h | Librispeech + Libri-Light |
| wav2vec2-base-960h-lv60-self | 960h | Librispeech + Libri-Light + Self Training |
Clone the project,
git clone https://github.com/ranchlai/wav2vec2.paddle
cd wav2vec2.paddleRun the speech recognition test with your audio file,
python test.py --device "gpu:0" --audio "LJ001-0186.wav" --config "wav2vec2-large-960h-lv60"If successful, you will see output like this,
pred==> position of our society that a work of utility might be also a work of art if we cared to make it so
If you do not have gpu or run out of gpu memory, try cpu:
python test.py --device "cpu" --audio "LJ001-0186.wav" --config "wav2vec2-large-960h-lv60"[1] Baevski, Alexei, et al. “Wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations.” Advances in Neural Information Processing Systems, vol. 33, 2020, pp. 12449–12460.