Abstract

This document is the description of Far-field Speaker Verification Challenge (FFSVC) 2022. The success of FFSVC2020 [1] indicates that more and more researchers are paying attention to the far-field speaker verification task. In this year, the challenge is still focus on the far-field speaker verification scenario and provided a new far-field development and test set collected by real speakers in complex environments with multiple scenarios, eg, text-dependent, text-independent, cross-channel enroll/test, etc. In addition, in real scenario, speech data is not alway labeled especially for far-field data, which is hard to accurate labeled with close-talking pre-trained model. Therefore, a new focus of this year is cross-language self-supervised/semisupervised learning, where participants are allowed to generate the pseudo-label for the train/dev set without using speaker label of the FFSVC2020 dataset (in Mandarin) by close-talking model trained by VoxCeleb1&2 (mostly in English) to fine-tune the model.
In contrast to FFSVC2020 tasks, this challenge is focus on the single-channel scenario, which means that both the enrollment and test audio are single-channel data. In addition, considering the real application scenario, a list of trials in this year will more challenge than FFSVC2020. The trial pairs will considering more hard cases, eg same gender, cross-domain, crosschannel, cross-time, etc.

Date

2022

Authors

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