This paper describes ID R&D team submission to the text-independent task of the Short-duration Speaker Verification (SdSV) Challenge 2021. The top performed system is a fusion of 9 Convolutional Neural Networks based on the ResNet architecture. Experiments’ results of optimal NN architecture search are shown. We also present and investigate the subnetwork approach to solve the auxiliary tasks such as gender or language detection. Verification scores refinement step using quality measurements of a trial pair allowed to further minimize the target metrics. A comparative analysis of all systems used in the fusion has been provided on the VoxCeleb-1 test set, SdSV-2021 development and evaluation sets. The final submission achieves 0.69% EER and 0.0319 minDCF on the challenge evaluation set.