Twinkle-ASR

Repo for processing data to submit to the VLSP2025 ASR/SER task.

1. Set up environment and install dependencies

1.1 Install uv and create UV environment

This is an extremely fast Python package and project manager, written in Rust. Refer to the uv link to reach it.

curl -LsSf https://astral.sh/uv/install.sh | sh

Creating a virtual environment using UV:

uv venv VLSP2025 --python python3.10
source VLSP2025/bin/activate

1.2 Install required dependencies and dependencies

To install the required libraries and frameworks, follow the steps below:

uv pip install -r requirements.txt
uv pip install -r requirements_2.txt
uv pip install -r requirements_VLSP.txt

Note: Due to conflicts between libraries, it's not possible to include them all in a single .txt file.

1.3 Install beam search to improve ASR accuracy

Refer to this link if it is still accessible.
Before proceeding, make sure to update the two variables NEMO_PATH and export PATH to the correct paths.
Then run sudo sh Install_beam.sh to install kenLM decoder for beamsearch.