license: apache-2.0
Pre-training wav2vec2 models for Welsh speech recognition
At the moment, the best Welsh speech recognition models are achieved from fine-tuning https://huggingface.co/facebook/wav2vec2-large-xlsr-53 and https://huggingface.co/facebook/wav2vec2-xls-r-1b models by Facebook/Meta AI.
This model is experimental in investigating pretraining better models with more Welsh language speech that could lower WER scores even further in subsequently fine-tuned models. The work draws heavily on resources and documentation from the HuggingFace examples:
https://github.com/huggingface/transformers/tree/main/examples/pytorch/speech-pretraining
This initial base model has been pre-trained with scripts at
https://github.com/techiaith/docker-wav2vec2-cy/tree/main/train/pre-train
using English speech from LibriSpeech's minimal subsets (validation and test), and 184 hours and 47 minutes of Welsh speech from various playlists on YouTube. The script build_youtube_playlists_corpus.sh lists the playlists used.
Until we have collected thousands of hours of Welsh speech, rather than hundreds, the WER scores, after fine-tuning, will remain very high. The following WERs are from tests on a Welsh Common Voice test set as well a second set of YouTube videos with corrected transcriptions.
| Test Set | WER | CER | WER (+LM) | CER (+LM) |
|---|---|---|---|---|
| CV CY 10 | 94.83 | 85.55 | 92.31 | 82.25 |
| YouTube | 95.43 | 90.26 | 93.60 | 89.33 |