Speaker verification project

by Andrii Shevtsov, Artur Shevtsov and Maksym Palamariuk

This is a realization of speaker verification system using pure numpy (PLDA technique) and its comparison with more advanced methods, such as Siamese Convolutional Neural Networks.

Folder structure

Folder structure is as follows:

• data folder is folder to store the dataset.
• features folder is for feature extraction. Those features are mainly used in PLDA realization.
• neural folder is for neural network-based Speaker verification system realization.
• plda folder is for realization of the PLDA algorithm and its enhancements.
• utils folder is for other project-related stuff that is general or inconvinient to put into previous folders.

Environment

We are using Python's narive virtual environments along with pip package manager and requirements.txt files.

To run the project, create a virtual environment via python<version> -m venv .venv.

Then, activate the environment:

• On Linux/Mac: source .venv/bin/activate.
• On Windows: .venv/Scripts/activate.bat.

And install requirements.txt:

pip install -r requirements.txt

If you want to use neural net, you need also to install neural/requirements.txt.

Dataset

To obtain the dataset, use the following steps:

1. Create a data folder in the project.

2. Run all cells in ./utils/download_data.ipynb