Voice Privacy Challenge

Welcome to the Voice Privacy Challenge! Your task is to develop a model that anonymizes audio while preserving intelligibility and naturalness. This repository provides the necessary setup, evaluation script, and rules for participation.

---

📂 Repository Structure

evaluation_data/       # Directory containing enrollment and trial audio data
│── Enrollment/        # Speaker audio files for enrollment
│   ├── speaker1/      # Directory for Speaker 1
│   │   ├── 1272-128104-0000.wav  # Original enrollment utterance
│   │   ├── ...
│   │   └── anonymized/            # Anonymized versions of the above audio files (Will be automatically created based on your anonymization algorithm when the evaluation script is run)
│   │       ├── anon_1272-128104-0000.wav
│   │       ├── ...
│   ├── speaker2/
│   ├── speaker3/
│   ├── speaker4/
│   └── ...
│
│── Trial/             # Speaker audio files for testing (trial phase)
│   ├── speaker1/
│   │   ├── 1272-128104-0003.wav  # Trial utterances (different from enrollment)
│   │   ├── ...
│   │   └── anonymized/           # Anonymized versions of the above audio files (Will be automatically created based on your anonymization algorithm when the evaluation script is run)
│   │       ├── anon_1272-128104-0003.wav
│   │       ├── ...
│   ├── speaker2/
│   ├── speaker3/
│   ├── speaker4/
│   └── ...
│
parameters/            # Directory to store model parameters (participants should add their own)
evaluation.py          # DO NOT MODIFY - Evaluates your model and generates results.csv
model.py               # MODIFY - Implement your anonymization model here
README.md              # This file - contains all competition instructions
requirements.txt       # MODIFY - List your dependencies here
run.sh                 # DO NOT MODIFY - Runs the evaluation script

---

🗂 Understanding Enrollment and Trial Data

In this challenge, participants work with enrollment and trial utterances, which follow a structure similar to speaker verification tasks.

• Enrollment Utterances (Stored in Enrollment/):

  • These are speech recordings associated with a particular speaker.
  • Each speaker has multiple enrollment utterances, which serve as reference data.
  • The anonymization system must ensure that any transformed enrollment utterance still preserves the necessary speech characteristics, except for the speaker's identity.

• Trial Utterances (Stored in Trial/):

  • These are new speech recordings from the same speakers but contain different utterances.
  • These utterances are anonymized and later compared against enrollment utterances.
  • The anonymization system must ensure that the same speaker's trial utterances still match their anonymized enrollment utterances while preventing identification of the original speaker.

🔑 Key Properties

• Each speaker in Enrollment and Trial is the same, meaning speaker1 in Enrollment/ is the same as speaker1 in Trial/, but their audio files differ.
• The anonymized versions of a speaker’s trial utterances must match the anonymized version of their enrollment utterances, maintaining consistency in the "pseudo-speaker" identity.
• The anonymization system should not alter linguistic content but should make it impossible to link the anonymized voice back to the original speaker.

---

🚀 Getting Started

1️⃣ Fork the Repository

Before cloning, you need to fork this repository to your own GitHub account. Follow these steps:

1. Navigate to the repository on GitHub.
2. In the top-right corner, click the Fork button.
3. This creates a copy of the repository under your GitHub account.

---

2️⃣ Clone Your Forked Repository

Once you've forked the repository, clone it to your local machine:

# Replace <YOUR_GITHUB_USERNAME> with your actual GitHub username
git clone https://github.com/<YOUR_GITHUB_USERNAME>/VPC25.git
cd VPC25

This ensures you're working on your own version of the repository while still being able to pull updates from the original source.

---

3️⃣ Set Up Your Environment

This project requires Python 3.12. Ensure you have it installed before proceeding.

Check your Python version:

python3 --version

or on Windows (PowerShell):

python --version

If you don't have Python 3.12, download it from python.org.

Install FFmpeg (Required for Audio Processing)

To process audio files, FFmpeg must be installed. Follow these steps based on your system:

Linux

sudo apt update && sudo apt install ffmpeg

macOS

brew install ffmpeg

Windows

1. Download FFmpeg from ffmpeg.org (recommended: Windows build from gyan.dev).
2. Extract it to a folder (e.g., C:\ffmpeg).
3. Add C:\ffmpeg\bin to your system PATH to make FFmpeg accessible from the command line.
4. Verify installation by running:

   ffmpeg -version

Create a Virtual Environment

These instructions should be followed inside the VPC25/ folder exactly as written. Do not modify the command examples, including the virtual environment name.

Linux/macOS

python3.12 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Windows (PowerShell)

python -m venv .venv
.venv\Scripts\Activate
pip install -r requirements.txt

This ensures all dependencies are installed inside an isolated environment.

Activate the Virtual Environment Each Time You Work on the Project

Each time you start working on the project, you should activate the virtual environment:

Linux/macOS

source .venv/bin/activate

Windows (PowerShell)

.venv\Scripts\Activate

For more details on virtual environments in Python, refer to:

• Python Virtual Environments Documentation
• Real Python Guide to Virtual Environments

---

4️⃣ Implement Your Model

• Modify model.py to implement your anonymization approach.
• Store any necessary model parameters in the parameters/ directory.
• Add any additional dependencies to requirements.txt.

⚠️ DO NOT modify:

• evaluation.py
• run.sh

---

5️⃣ Add Your Source Audio Files

• Place your test audio files inside evaluation_data/.
• The evaluation script will process these files automatically.

---

6️⃣ Run the Evaluation

To test your model, execute:

bash run.sh

This will:

1. Set up and activate the virtual environment (if not already done).
2. Ensure dependencies are installed.
3. Process the source audio.
4. Generate anonymized audio files.
5. Output evaluation results to results.csv.

Important:

• Windows users must use Git Bash to run this command, as PowerShell and Command Prompt do not support shell scripts properly.
• Windows and macOS users might need to run run.sh with administrator privileges to avoid permission issues with symbolic links.

---

📊 Evaluation Metrics

The evaluation script will measure:

• Equal Error Rate (EER): This metric, derived from an Automatic Speaker Verification (ASV) system, measures the system's ability to differentiate between speech from the same speaker and different speakers. A higher EER indicates better privacy protection, as it means the system is less likely to correctly identify the speaker.
• Word Error Rate (WER): This metric is calculated using an Automatic Speech Recognition (ASR) system and measures how well the anonymized speech preserves linguistic content. A lower WER indicates better utility, meaning the anonymized speech is still easily understood by the ASR system.
• Processing time: Measure the effeciency of the anonymization algorithm.

Results are stored in results.csv.

---

📜 Rules & Guidelines

✅ You MUST:

• Implement your model in model.py.
• List dependencies in requirements.txt.
• Store model parameters in parameters/.
• Run evaluation using run.sh.

❌ You MUST NOT:

• Delete or modify evaluation.py or run.sh.
• Remove or alter existing directories.

Good luck! 🚀🎧