[Cherry-picked 0.10] Add pretrained weights from wav2vec2.0 and XLSR papers #1827
Conversation
| ) | ||
| WAV2VEC2_ASR_BASE_10M.__doc__ = """Build "base" wav2vec2 model with an extra linear module | ||
| Pre-trained on 960 hours of *LibriSpeech* [:footcite:`7178964`] dataset, and |
There was a problem hiding this comment.
Does this correspond to the Wav2Vec 2.0 Large | 10 minutes entry in the table? If so, should it be fine-tuned on LibriSpeech instead of Libri-Light?
There was a problem hiding this comment.
Libri-Light is a subset of LibriSpeech, so both description is correct, but Libri-Light is more accurate.
Here is the description from the wav2vec 2.0 paper.
We fine-tune on five labeled data settings: 960 hours of transcribed Librispeech, the train-clean-100 subset comprising 100 hours (100 hours labeled), as well as the Libri-light limited resource training subsets originally extracted from Librispeech, these are train-10h (10 hours labeled), train-1h (1 hour labeled), train-10min (10 min labeled).
|
|
||
| WAV2VEC2_ASR_BASE_100H.__doc__ = """Build "base" wav2vec2 model with an extra linear module | ||
| Pre-trained and fine-tuned for ASR on 960 hours of |
There was a problem hiding this comment.
I think this is switched with the WAV2VEC2_ASR_BASE_960H doc below
There was a problem hiding this comment.
Good catch! Thank you!
| ) | ||
| WAV2VEC2_ASR_LARGE_LV60K_10M.__doc__ = """Build "large-lv60k" wav2vec2 model with an extra linear module | ||
| Pre-trained on 60,000 hours of *Libri-Light* [:footcite:`librilight`] dataset, and |
There was a problem hiding this comment.
From this table and your WAV2VEC2_ASR_LARGE_LV60K_100H doc below, I think this should be fine-tuned on LibriSpeech instead of Libri-Light
There was a problem hiding this comment.
Thanks for spotting the error. I looked at the paper again and it turned out that LibriVox is the correct one.
The following is the relationship between these datasets.
LibriVox: 60,000 hours audioLibriSpeech: 960 hours audio + transcript, subset ofLibriVoxLibriLight(Limited Resource Training Set): subset ofLibriSpeechtraining subset
Add pretrained weights from https://github.com/pytorch/fairseq/tree/main/examples/wav2vec#pre-trained-models - Wav2Vec 2.0 Base / Large / Large (LV-60) - XLSR-53
Co-authored-by: Caroline Chen <[email protected]>
Add pretrained weights from https://github.com/pytorch/fairseq/tree/main/examples/wav2vec#pre-trained-models - Wav2Vec 2.0 Base / Large / Large (LV-60) - XLSR-53
mthrok
changed the title
4 participants
Add pretrained weights from https://github.com/pytorch/fairseq/tree/main/examples/wav2vec#pre-trained-models