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.

Consider a weighted, undirected graph cellularly embedded on a topological surface. The function assigning to each free homotopy class of closed curves the length of a shortest cycle within this homotopy class is called the marked length spectrum. The (unmarked) length spectrum is obtained by just listing the length values of the marked length spectrum in increasing order. In this paper, we describe algorithms for computing the (un)marked length spectra of graphs embedded on the torus. More specifically, we preprocess a weighted graph of complexity $n$ in time $O(n^2 \log \log n)$ so that, given a cycle with $\ell$ edges representing a free homotopy class, the length of a shortest homotopic cycle can be computed in $O(\ell+\log n)$ time. Moreover, given any positive integer $k$, the first $k$ values of its unmarked length spectrum can be computed in time $O(k \log n)$. Our algorithms are based on a correspondence between weighted graphs on the torus and polyhedral norms. In particular, we give a weight independent bound on the complexity of the unit ball of such norms. As an immediate consequence we can decide if two embedded weighted graphs have the same marked spectrum in polynomial time. We also consider the problem of comparing the unmarked spectra and provide a polynomial time algorithm in the unweighted case and a randomized polynomial time algorithm otherwise.

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 and hosted on GDrive1. Baseline code and pre-trained models are available at GitHub2.