An ONNX model for speech recognition of the Ukrainian language

Overview

This repository contains an ONNX model for speech recognition of the Ukrainian language exported from wav2vec2 1b model.

If you want to export own ONNX model, follow this Google Colab.

Installation

Download onnx-uk-1b.zip (3.33 GB) file and unpack it in the repository folder.

There's the quantized version of the model (~300 MB), link: onnx-uk-300m-quantized.onnx .

Note: replace vocab.json by vocab_300m.json file if you will use the quantized version.

Install Python dependencies:

pip install onnxruntime numpy scipy

Running

python recognize.py

Demo

The quantized version the ONNX model is running on a Jetson Nano computer:

Jetson Nano is charged by a Power Bank:

Links

• Other Speech-to-Text implementations: https://github.com/egorsmkv/speech-recognition-uk
• Telegram community about speech recognition of the Ukrainian language: https://t.me/speech_recognition_uk

Tensorflow

If you like to use Tensorflow, then there's a SavedModel - tf_uk_300m_model.zip

Use tf_predict.py to recognize files with the Tensorflow model.