robinhad / voice-recognition-ua

Training scripts for Speech-To-Text models for Ukrainian language

🔍 Guide to my GitHub repositories:

Repository	Project description
ukrainian-tts 📢🤖	Ukrainian text-to-speech
kruk 🤖	Ukrainian instruction-tuned language models and datasets
qirimtatar-tts 📢🤖	Crimean Tatar text-to-speech
voice-recognition-ua 🤖🎤	Speech-to-Text training scripts for Ukrainian WER 12,22%
ukrainian-ai 🤖	End-to-End voice bot, that you can talk to in Ukrainian
ukrainian-qa ❓	Ukrainian Question Answering models (Extractive and Generative)
photoscan 🏛️👀	Application for obtaining properties (such as angle and position) of historical photo
uk-voyager-audio-dataset	Toy dataset that contains Ukrainian recording from Voyager Golden Disc

Link to HuggingFace profile: https://huggingface.co/robinhad

Link to LinkedIn profile: https://www.linkedin.com/in/yurii-paniv-aa969b151/

voice-recognition-ua's Issues

Update guides for training and dataset import

Limit audio duration to 10 seconds

DeepSpeech model has a default batch size of 10 seconds.

Train scorer for words

This is a tracking issue for scorer support.

For training your own scorer please use Mozilla guide.

Checklist:

• Gather Ukrainian texts (at least 4 GB as English model) (2,5 GB Wikipedia + 700 MB of texts)
• Train scorer
• Use scorer in client.py

Train new model with simplified alphabet

Include only significant characters in alphabet.txt, such as [а-я] + ['].

Bundle checkpoints with newer releases

Thank you for open-sourcing your work, @robinhad!

In older releases [1] you release your training checkpoints, but in newer releases [2] you don't. It would be great to release the checkpoints in newer releases as well, so others can extend the models as they get better.

Thanks again for open-sourcing this work!

Add resampling to demo

Train/fine-tune Facebook model

See link here https://huggingface.co/facebook/wav2vec2-large-xlsr-53

https://huggingface.co/anton-l/wav2vec2-large-xlsr-53-ukrainian/tree/main