Wav2vec resources

Wav2Vec tools:

1. Huggingface

• Advantage: compatible with other huggingface layers
• Disadvantage: currently not supporting language model for decoding, will be updated someday.
• Good resource to take a look at: Fine-Tune Wav2Vec for English ASR with Huggingface Transformers

• If you want to follow the provided resource with your custom dataset, take a look at this notebook.

2. Fairseq

• Advantage: support language model for decoding. YAML file for all hyperparameters.
• Disadvantage: requires large computation power (original study used 24 gpu, and it shows OOM error when running on Google Colab)
• TODO: Upload custom scripts that wrote to preprocess the data structure.
• Install flashlight:

• How to fix CANNOT FIND FFTW3LibraryDependency error?: github issue
• export MKLROOT path: github issue

3. SpeechBrain

• Advantage: compatible with huggingface. Support language model. Good documentation
• Disadvantage: requires python >= 3.8. need to understand the structure of yaml files and custom functions.
• TODO:

• Wav2Vec2.0 with huggingface and fairseq using SpeechBrain HERE
• Pretrain and finetune the model using huggingface HERE