spacejake / Voice-Privacy-Challenge-2022

Baseline Recipe for VoicePrivacy Challenge 2022: anonymization systems and evaluation software

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Recipe for VoicePrivacy Challenge 2022

Please visit the challenge website for more information about the Challenge.

Install

  1. git clone --recurse-submodules https://github.com/Voice-Privacy-Challenge/Voice-Privacy-Challenge-2022.git
  2. ./install.sh

Running the recipe

The recipe uses the pre-trained models of anonymization. To run the baseline system with evaluation:

  1. cd baseline
  2. run ./run.sh. In run.sh, to download models and data the user will be requested the password which is provided during the Challenge registration.

General information

For more details about the baseline and data, please see The VoicePrivacy 2022 Challenge Evaluation Plan

For the latest updates in the baseline and evaluation scripts, please visit News and updates page

To participate in the VoicePrivacy 2022 Challenge and get access to evaluation datasets and models, please send an email to organisers@lists.voiceprivacychallenge.org with “VoicePrivacy-2022 registration" as the subject line. The mail body should include:

  • (i) the name of the team;
  • (ii) the name of the contact person;
  • (iii) their affiliation;
  • (iv) their country;
  • (v) their status (academic/nonacademic).

Data

Training data

The dataset for anonymization system traing consists of subsets from the following corpora*:

*only specified subsets of these corpora can be used for training.

Development and evaluation data

  • VCTK - subsets vctk_dev and vctk_test are download from server in run.sh
  • LibriSpeech - subsets libri_dev and libri_test are download from server in run.sh

Baselines

Baseline B1.a: Anonymization using x-vectors and neural waveform models

This is the same baseline as the primary baseline for the VoicePrivacy-2020. In config.sh parameters: baseline_type=baseline-1 tts_type=am_nsf_old

Models

The baseline B1.b system uses several independent models:

  1. ASR acoustic model to extract BN features (1_asr_am) - trained on LibriSpeech-train-clean-100 and LibriSpeech-train-other-500
  2. X-vector extractor (2_xvect_extr) - trained on VoxCeleb 1 & 2.
  3. Speech synthesis (SS) acoustic model (3_ss_am) - trained on LibriTTS-train-clean-100.
  4. Neural source filter (NSF) model (4_nsf) - trained on LibriTTS-train-clean-100.

All the pretrained models are provided as part of this baseline (downloaded by ./baseline/local/download_models.sh)

Baseline B1.b: Anonymization using x-vectors and neural waveform models (HiFi-GAN + NSF)

The main difference wrt to B1.a is in the speech synthesis component of the anonymization system, B1.b directly converts BN features, F0, and x-vector using an NSF model.

Models

The baseline B1.b system uses several independent models:

  1. ASR acoustic model to extract BN features (1_asr_am) - trained on LibriSpeech-train-clean-100 and LibriSpeech-train-other-500
  2. X-vector extractor (2_xvect_extr) - trained on VoxCeleb 1 & 2.
  3. Speech synthesis model HiFi-GAN + NSF (5_joint_tts_nsf_hifigan) - trained on LibriTTS-train-clean-100.

In config.sh parameters: baseline_type=baseline-1 tts_type=joint_nsf_hifigan

Baseline B2: Anonymization using McAdams coefficient (randomized version)

This ia a randomized version of the McAdams algorithm, where the McAdams coefficient is sampled for each source speaker in the evaluation set from a uniform distribution in the interval [mc_coeff_min, mc_coeff_max]

In config.sh parameters:
baseline_type=baseline-2 mc_coeff_min=0.5 mc_coeff_max=0.9

It does not require any training data and is based upon simple signal processing techniques using the McAdams coefficient.

Results

The result file with all the metrics and all datasets for submission will be generated in:

  • Summary results: ./baseline/exp/results-<date>-<time>/results_summary.txt
  • Additional metrics obtained using ASR_eval and ASV_eval trained on original data: ./baseline/exp/results-<date>-<time>.orig/results_summary.txt
  • Additional metrics obtained using ASR_eval^anon and ASV_eval^anon trained on anonymized data: ./baseline/exp/results-<date>-<time>/results_summary.txt

Please see

for the evalation and development data sets.

Organizers (in alphabetical order)

  • Jean-François Bonastre - University of Avignon - LIA, France
  • Nicholas Evans - EURECOM, France
  • Pierre Champion - Inria, France
  • Xiaoxiao Miao - NII, Japan
  • Hubert Nourtel - Inria, France
  • Natalia Tomashenko - University of Avignon - LIA, France
  • Massimiliano Todisco - EURECOM, France
  • Emmanuel Vincent - Inria, France
  • Xin Wang - NII, Japan
  • Junichi Yamagishi - NII, Japan and University of Edinburgh, UK

Contact: organisers@lists.voiceprivacychallenge.org

Acknowledgements

This work was supported in part by the French National Research Agency under project DEEP-PRIVACY (ANR-18- CE23-0018) and by the European Union’s Horizon 2020 Research and Innovation Program under Grant Agreement No. 825081 COMPRISE (https://www.compriseh2020.eu/), and jointly by the French National Research Agency and the Japan Science and Technology Agency under project VoicePersonae.

License

Copyright (C) 2021

This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with this program. If not, see https://www.gnu.org/licenses/.


References

article{vpc2022,
  title={The {VoicePrivacy} 2022 {Challenge} Evaluation Plan},
  author={Tomashenko, Natalia and Wang, Xin and Miao, Xiaoxiao and Nourtel, Hubert and Champion, Pierre and Todisco, Massimiliano and Vincent, Emmanuel and Evans, Nicholas and Yamagishi, Junichi and Bonastre, Jean-François
},
  url={https://www.voiceprivacychallenge.org/docs/VoicePrivacy_2022_Eval_Plan_v1.0.pdf},
  year={2022}
}

Anonymization metrics

VoicePrivacy 2022 primary privacy metric:

  • Equal error rate (EER)

Other metrics:

About

Baseline Recipe for VoicePrivacy Challenge 2022: anonymization systems and evaluation software


Languages

Language:Python 79.0%Language:Shell 20.9%Language:CMake 0.1%