Nikita Torgashov

In this paper, we present Reshape Dimensions Network (ReDimNet), a novel neural network architecture for extracting utterance-level speaker representations. Our approach leverages dimensionality reshaping of 2D feature maps to 1D signal representation and vice versa, enabling the joint usage of 1D and 2D blocks. We propose an original network topology that preserves the volume of channel-timestep-frequency outputs of 1D and 2D blocks, facilitating efficient residual feature maps aggregation. Moreover, ReDimNet is efficiently scalable, and we introduce a range of model sizes, varying from 1 to 15 M parameters and from 0.5 to 20 GMACs. Our experimental results demonstrate that ReDimNet achieves state-of-the-art performance in speaker recognition while reducing computational complexity and the number of model parameters.

The objective of this paper is to advance the development of technologies in the fields of speaker recognition and speaker identification by introducing a large labeled audio database VoxTube collected from the open-source media. We propose a fully automated unsupervised approach for audio labeling that requires any pre-trained speaker recognition model. Collected with this approach from videos with CC BY license the VoxTube dataset contains more than 5.000 speakers with more than 4 million utterances pronounced in more than 10 languages. In our paper we show the VoxTube’s high generalization ability across multiple domains by evaluating the accuracy metrics on various speaker recognition benchmarks. We also show how well this dataset complements an already existing VoxCeleb2 dataset.

This paper describes the ID R&D team submission to the Spoofing Aware Speaker Verification (SASV) challenge. Firstly, we present an approach that utilizes automatic speaker verification (ASV) system together with countermeasures (CM) subsystem in a single computational graph, called an antispoofing subnetwork. Subnetwork is a small network operating on feature maps from larger parent network, in this case trained for a speaker verification task. While requiring a small number of additionally trained parameters, subnetwork approach showed great performance in spoofing attacks detection task. Secondly, we cover training strategies for independently trained ASV and CM systems. In addition, we present a SASV-EER optimization approach using a fusion of multiple systems outputs and quality measurement functions (QMFs). Our best fusion achieves 0.136% EER on SASV-2022 evaluation set, while the smallest single-model system with 11.6 M parameters achieves 0.223% EER.

The latest research in the field of voice anti-spoofing (VAS) shows that deep neural networks (DNN) outperform classic approaches like GMM in the task of presentation attack detection. However, DNNs require a lot of data to converge, and still lack generalization ability. In order to foster the progress of neural network systems, we introduce a Large Replay Parallel Dataset (LRPD) aimed for a detection of replay attacks. LRPD contains more than 1M utterances collected by 19 recording devices in 17 various environments. We also provide an example training pipeline in PyTorch [1] and a baseline system, that achieves 0.28% Equal Error Rate (EER) on evaluation subset of LRPD and 11.91% EER on publicly available ASVpoof 2017 [2] eval set. These results show that model trained with LRPD dataset has a consistent performance on the fully unknown conditions. Our dataset is free for research purposes.