librivox-errors

For project in course DT2112 VT24, "Detecting and correcting speech errors produced on LibriVox audiobooks".

Phoneme model and fine-tuning with 5-gram

We used a wav2vec2 model and feed it with a 5-gram model provided to us by our supervisor. https://www.kaggle.com/code/julwan/dubliners-wav2vec2-xls-r-300m-timit-phoneme

Error metrics

Word error rate

https://www.kaggle.com/code/julwan/wer-dubliners

Distance alignment confusion matrix

https://www.kaggle.com/code/julwan/distance-alignment-confusion-matrix-dubliners

Over characters and phones between the two different accents.

Report

The report can be found on https://www.overleaf.com/read/qhtvzkrjthms#83b9f5

Credit

Thanks to our supervisor Jim O'Regan for helping us through the project and providing us with the materials necessary to work on this project.

Team

• Julia Wang
• Kevin Wenström
• Peter Cady