Download data from https://huggingface.co/datasets/ProgramComputer/voxceleb and place it inside the data folder.

DOWNLOAD ONLY THE VoxCeleb1 dataset (the smaller one, ~30-40 GB).

The data folder should contain two directories:

• vox1_dev_wav
• vox1_test_wav

Each of those folders should have the following structure:

• wav folder ↓
  • many directories named like idXYZPQ ↓
    • many directories with random names,
      example: 1zcIwhmdeo4 ↓
      • *.wav files

The data folder should also contain a vox1_vox1_meta.csv. It's tracked by git, so it should be there by default.

Project stages are contained in jupyter notebooks in the src/notebooks directory in order to facilitate quick edits and live data inspection.

Stages:

1. annotate_audio_files - checks audio files from the VoxCeleb1 dataset and creates a annotations_with_metadata.json file, which contains all the relevant information for each audio recording - it's path relative to the project root, user_id, train/test split etc.
2. create_raw_embeddings - given an embedding SpeechBrain model name and an annotation file containing audio paths relative to the project root creates embeddings for all audio recordings and saves them using the pickle format.

Stage dependencies graph:

graph LR
    A[1. annotate_audio_files] --> B[2. create_raw_embeddings]