VoicePAT: Voice Privacy Anonymization Toolkit

Note: This repository and its documentation are still under construction but can already be used for both anonymization and evaluation. We welcome all contributions to introduce more generation methods or evaluation metrics to the VoicePAT framework. If you are interested in contributing, please leave comments on a GitHub issue.

VoicePAT is a toolkit for speaker anonymization research. It is based on the framework(s) by the VoicePrivacy Challenges but contains the following improvements:

• It consists of two separate procedures for anonymization and evaluation. This means that the generation of anonymized speech is independent of the evaluation of anonymization systems. Both processes do not need to be executed in the same run or with the same settings. Of course, you need to perform the anonymization of evaluation data with one system before you can evaluate it but this could have happened at an earlier time and with an external codebase.
• Anonymization and evaluation procedures are structured as pipelines consisting of separate modules. Each module may have a selection of different models or algorithm to fulfill its role. The settings for each procedure / pipeline are defined exclusively in configuration files. See the Usage section below for more information.
• Evaluation models have been exchanged by models based on SpeechBrain and ESPnet which are more powerful than the previous Kaldi-based models. Furthermore, we added new techniques to make evaluation significantly more efficient.
• The framework is written in Python, making it easy to include and adapt other Python-based models, e.g., using PyTorch. When using the framework, you do not need in-depth knowledge about anything outside the Python realm (Disclaimer: While being written in Python, the ASR evaluation is currently included with an ESPnet-based model which in turn is based on Kaldi. However, you do not need to modify that part of the code for using or changing the ASR model and ESPnet is currently working on a Kaldi-free version.)

Installation

Requires conda for environment management. Installation of mamba is also recommended for speeding up the environment-related tasks. Simply clone the repository and run the following commands, a conda environment will be generated in the project root folder and the pretrained models will be downloaded.

sudo apt install libespeak-ng   # alternatively use your own package manager
make install pretrained_models  

The datasets have to be downloaded via the VoicePrivacy Challenge framework. Once the download is complete, the .scp files need to be converted to the absolute path, because they are relative to the challenge folder. Use utils/relative_scp_to_abs.py for this purpose. Then simply point data_path in the YAML configurations to the data folder of the VoicePrivacy Challenge framework.

If you want to use the ESPnet-based ASR evaluation model, you additionally need to clone and install ESPNet and insert the link to it in evaluation/utility/asr/path.sh, e.g., MAIN_ROOT=~/espnet.

Usage

For using the toolkit with the existing methods, you can use the configuration files in configs. You can also add more modules and models to the code and create your own config by using the existing ones as template. The configuration files use HyperPyYAML syntax, for which a useful reference is available here.

Anonymization

The framework currently contains only one pipeline and config for anonymization, anon_ims_sttts_pc.yaml. If you are using this config, you need to modify at least the following entries:

data_dir:    # path to original data in Kaldi-format for anonymization
results_dir: # path to location for all (intermediate) results of the anonymization
models_dir:  # path to models location

Running an anonymization pipeline is done like this:

python run_anonymization.py --config anon_ims_sttts_pc.yaml --gpu_ids 0,1 --force_compute

This will perform all computations that support parallel computing on the gpus with ID 0 and 1, and on GPU 0 otherwise. If no gpu_ids are specified, it will run only on GPU 0 or CPU, depending on whether cuda is available. --force_compute causes all previous computations to be run again. In most cases, you can delete that flag from the command to speed up the anonymization.

Pretrained models for this anonymization can be found at https://github. com/DigitalPhonetics/speaker-anonymization/releases/tag/v2.0 and earlier releases.

Evaluation

All other config files in configs can be used for evaluation with different settings. In these configs, you need to adapt at least

eval_data_dir: path to anonymized evaluation data in Kaldi-format
asr/libri_dir: path to original LibriSpeech dataset

Running an evaluation pipeline is done like this:

python run_evaluation.py --config eval_pre_ecapa_cos.yaml --gpu_ids 1,2,3

making the GPUs with IDs 1, 2 and 3 available to the process. If no GPU is specified, it will default to CUDA:0 or use all GPUs if cuda is available, or run on CPU otherwise.

Pretrained evaluation models can be found in release v1.

Acknowledgements

Several parts of this toolkit are based on or use code from external sources, i.e.,

• VoicePrivacy Challenge 2022, ESPnet, SpeechBrain for evaluation
• the GAN-based anonymization system by IMS (University of Stuttgart) for anonymization

See the READMEs for anonymization and evaluation for more information.