Huggingfaces released a model for facebooks fairseqs Wave2Vec2 for low resource speech recognition(sound 2 text). The following was my attempt to finetune and train on the malay language and there are several other repositories doing this but i thought why not try and fine tune for the sprint.
The langauge i was attempting is in bahasa malaysia (Malay). The other is a special form of english here in Malaysia (Manglish). So far the attempted finetuning method is Malay. Will attempt to gather more data on manglish
- (30/3/2021) : Training Failed, Need to debug
- (30/3/2021: 4:00pm) : Fix logging step and added evaluation step, change new data inputs and fix data gathering methods
- (30/3/2021): Word Error Rate is 1 (need to debug reasons for high error rate)
- Gather Data
- EDA
- Preprocesing
- Train Wave2Vec2 model
- Inference Script
- Evaluate model
- Upload model to huggingface hub