Voice Anonymization Research Repository

This repository contains exploratory research code for voice anonymization and related audio processing tasks. It is organized for iterative experimentation rather than production-style packaging.

How this repo is meant to be used

• Start from reproducible scripts to prepare data and train models.
• Use notebooks for analysis/inspection after data and checkpoints are generated.
• Expect structure and files to evolve as research questions change.
• Prefer reproducible checkpoints and notes over strict software release workflows.

Repository structure

.
|-- anonymization_pipeline/   # Core anonymization pipeline components
|-- notebooks/                # Analysis and experiment inspection
|-- scripts/                  # Data prep and training entrypoints
|-- source_separation/        # Source separation experiments and related code
|-- voice_blurring/           # Voice blurring methods and prototypes
`-- requirements.txt          # Minimal Python dependencies

Quick start

1. Create and activate a Python environment.
2. Install dependencies:

 pip install -r requirements.txt

Preparing data

Use scripts/prepare_sonyc_vox_mixes.py to generate SONYC + VoxCeleb mixtures.

• Default inputs:
  • SONYC annotations: datasets/sonyc-v1-dataset/annotations.csv
  • SONYC audio root: datasets/sonyc-v1-dataset/
  • Vox root: datasets/voxceleb1-audio-wav-files-for-india-celebrity/
  • Vox metadata: datasets/voxceleb1-audio-wav-files-for-india-celebrity/vox1_meta.csv
• Output is written to the path passed with --out-dir, plus a manifest CSV.

Examples:

python scripts/prepare_sonyc_vox_mixes.py --mode train --out-dir data/mixes_train
python scripts/prepare_sonyc_vox_mixes.py --mode eval --out-dir data/mixes_eval --eval-snr high
python scripts/prepare_sonyc_vox_mixes.py --mode eval --out-dir data/mixes_eval --use-training-vox

Training U-Net

Train with scripts/train_unet.py. The script expects a CSV manifest with:

• mix_wav: path to a mixture WAV
• voice_wav: path to the corresponding target voice WAV

Example training command:

python scripts/train_unet.py --manifest data/pairs.csv --checkpoint-dir checkpoints/run1

Useful options:

python scripts/train_unet.py \
  --manifest data/pairs.csv \
  --checkpoint-dir checkpoints/run1 \
  --epochs 20 \
  --batch-size 8 \
  --lr 1e-3 \
  --device auto \
  --num-workers 4 \
  --seed 0

Output:

• A model checkpoint at checkpoints/<run>/unet_voice_sep.pt

References

• Voice anonymization reference listed in pdf-material/fonts.txt:
  • https://markcartwright.com/files/cohen-hadria2019voiceanonymization.pdf

Datasets used

• https://www.kaggle.com/datasets/gaurav41/voxceleb1-audio-wav-files-for-india-celebrity
• https://zenodo.org/records/3692954