Development of ABC systems for the 2021 edition of NIST Speaker Recognition evaluation

In this contribution, we provide a description of the ABC teamscollaborative efforts toward the development of speaker verificationsystems for the NIST Speaker Recognition Evaluation 2021 (NISTSRE2021).Cross-lingual and cross-dataset trials are the two mainchallenges introduced in the NIST-SRE2021. Submissions of ABCteam are the result of active collaboration of researchers from BUT,CRIM, Omilia and Innovatrics. We took part in all three close conditiontracks for audio-only, audio-visual and visual-only verificationtasks. Our audio-only systems follow deep speaker embeddings(e.g., x-vectors) with a subsequent PLDA scoring paradigm. As embeddingsextractor, we select some variants of residual neural network(ResNet), factored time delay neural network (FTDNN) andHybrid Neural Network (HNN) architectures. The HNN embeddingsextractor employs CNN, LSTM and TDNN networks and incorporatesa multi-level global-local statistics pooling method in orderto aggregate the speaker information within short time-span andutterance-level context. Our visual-only systems are based on pretrainedembeddings extractors employing some variants of ResNetand the scoring is based on cosine distance. When developing anaudio-visual system, we simply fuse the outputs of independent audioand visual systems. Our final submitted systems are obtainedby performing score level fusion of subsystems followed by scorecalibration.