希尔贝壳—专注于人工智能大数据和技术的创新

Interspeech 2022

FFSVC 2022

Far-field speaker verification challenge 2022

Welcome to FFSVC 2022! The success of FFSVC2020 indicates that more and more researchers are paying attention to the far-field speaker verification task. In this year, the challenge still focuses on the far-field speaker verification task and provides a new far-field development and test set collected by real speakers in complex environments under multiple conditions, e.g., text-dependent, text-independent, cross-channel enroll/test, etc. In addition, in-domain training speech data may be unlabeled, which is difficult to fine-tune the pre-trained model. Therefore, a new focus of this year is cross-language self-supervised / semi-supervised learning, where participants are allowed to use the unlabeled train and dev set of the in-domain FFSVC2020 dataset (in Mandarin) and the labeled out-of-domain VoxCeleb 1&2 dataset (mostly in English) to build the model.

Task Description

This year we focus on the far-field single-channel scenarios. There are two tasks in this challenge; both tasks are to determine whether two speech samples are from the same speaker：

Task 1. Fully supervised far-field speaker verification.

Task 2. Semi-supervised far-field speaker verification.

We define the task 1&2 as fixed training conditions that the participants can only use a fixed training set to build the speaker verification system. The fixed training set consists of the following two databases：

VoxCeleb 1&2.

FFSVC2020 dataset (Train and dev set).

FFSVC2020 supplementary set (new!).

Note: Please refer to this website to download VoxCeleb 1&2 dataset and this website to download FFSVC 2020 dataset if you do not have these two datasets. In addition, in this challenge, we release a supplementary set of FFSVC2020, which consists of the same devices data as FFSVC2022.

In task 1, participants can use both VoxCeleb1&2 and FFSVC20 datasets with speaker labels to train a far-field speaker verification system.

For task 2, in contrast to task1, participants cannot use the speaker labels of the FFSVC2020 dataset. In task 2, we encourage the participants to adopt self-supervised or semi-supervised methods to utilize the in-domain unlabeled data.

Using other speech datasets to train the system is forbidden, while participants are allowed to use public open-source non-speech dataset to perform data augmentation. The self-supervised pre-trained models, such as Wav2Vec and WavLM, cannot be used in this challenge.

Schedule

April 15th, 2022 : Releasing the FFSVC 2022 evaluation plan and starting the registration.
- April 20th, 2022 : Opening the submission system and releasing supplementary/dev/eval sets
- July 3th, 2022 : Deadline for registration.
July 10th, 2022 : Deadline for results submission.
- July 15th, 2022 : Deadline for system description submission
July 24th, 2022 : Deadline for workshop paper submission
Aug 20th, 2022 : Workshop paper acceptance notification
Sep 17th, 2022 : Interspeech 2022 Satellite Workshop

FFSVC 2020 Website

https://ffsvc.github.io/

Releasing the FFSVC 2022 evaluation plan and starting the registration

Prizes

Prizes will be awarded to top three winning teams of each task.

Organisers

Ming Li, Duke Kunshan University, China

Haizhou Li, National University of Singapore, Singapore

Shrikanth Narayanan, University of Southern California, USA

Hui Bu, AI Shell Foundation,China

Xiaoyi Qin, Duke Kunshan University, China

Yao Shi, Duke Kunshan University, China

Papers

Far-field Speaker Verification Challenge (FFSVC) 2022 : Challenge Evaluation Plan

Qin, Xiaoyi, Li, Ming, Bu, Hui, Narayanan, Shrikanth, and Li, Haizhou

2022

Evaluation Plan

ABS

HTML

BIB

PDF

This document is the description of Far-field Speaker Verification Challenge (FFSVC) 2022.

@article{FFSVC2022_Eval_Plan,
  abbr = {Evaluation Plan},
  bibtex_show = {true},
  title = {Far-field Speaker Verification Challenge (FFSVC) 2022 : Challenge Evaluation Plan},
  author = {Qin, Xiaoyi and Li, Ming and Bu, Hui and Narayanan, Shrikanth and Li, Haizhou},
  journal = {},
  html = {https://ffsvc.github.io/assets/pdf/ffsvc2022_plan_v2.pdf},
  pdf = {ffsvc2022_plan_v2.pdf},
  selected = {true},
  year = {2022}
}

Register

Since the challenge will be held on the Codalab platform, please create a Codalab account if you do not have one. We kindly request you to associate your account to an institutional e-mail. The organizing committee reserves the right to revoke your participant to the challenge otherwise, please read the evaluation plan carefully. Make sure to set the name of your team in the user's profile, or it will not be visible on the leaderboard.

The following is the Codalab links corresponding to each task:

Task 1. Fully supervised far-field speaker verification.
- Task 2. Semi-supervised far-field speaker verification.

Participants can register in one or two tasks. If your team participates in multiple tasks, we kindly request you to use the same user account to participate in all tasks.

Please note that any deliberate attempts to bypass the submission limit (for instance, by creating multiple accounts and using them to submit) will lead to automatic disqualification.

Submission

Participants are required to submit at least one valid score file for each participating task to the Codalab platform. Clicking on the “Submit/View Results” link under the “Participants” tab could submit your score file. You must submit your results in the form of a ZIP file, containing only one file named “scores.txt”. The file must be at the root of the ZIP file and the ZIP file should not contain any folders. The score files should be in UTF-8 format with one line per trial.