OK, a bug in the data preparation

Should be 'array': data[0], instead of 'array': data,

Published the testset for Ukrainian to HF:

https://huggingface.co/datasets/Yehor/cv10-uk-testset-clean

#ai #speech

Here is an integration of inference code with https://huggingface.co/Yehor/w2v-bert-uk

Colab: https://drive.google.com/file/d/1vBGPGQsLsKI9Yy_MNeageyjK6x-cLehC/view?usp=sharing

#ai #asr #speech

https://huggingface.co/Yehor/w2v-xls-r-uk requires about 1.2 GB of GPU memory with float16.

It gives following metrics (without an external LM):

Accuracy on words: 79.76%
Accuracy on chars: 96.36%

Around 300 million of parameters.

#asr #ai #speech

Added statistics over durations to the dataset.

Btw, Polars has useful describe method do calculate them as fast as lightning.

#ai #python

What happends when you use Whisper

#asr #ai #speech

openai/whisper-large-v3-turbo looks like does not recognize ї correctly

it's huggingface checkpoint

Evaluation results...

https://github.com/keyz

best gif in the github profile i have ever seen on gh

ok, testing official package to do inference

but all know what it is

#speech

the same

Official whisper package, seems like Hugging Face implementation has troubles with decoding...

#asr #ai

https://leaderboard.roboflow.com

#ai #cv

Use the following colabs to see how you can download this dataset in Python:

datasets: https://colab.research.google.com/drive/1qqnr5-WkaJi8iqHa_Pmlx7PbbXwXiimD?usp=sharing

polars: https://colab.research.google.com/drive/1upeXw3WbLjK37b1LetpM0HxFXDdOZqSK?usp=sharing

#ai #asr #speech

Moved VOSK models for Ukrainian to HF:

https://huggingface.co/Yehor/vosk-uk

#ai #speech #asr