A real-time trained system for robust speaker verification using relative space of anchor models

摘要: A real-time trained system for robust speaker verification is proposed. This was developed using a relative space of reference speakers, also referred to as anchor models. The training aspect the based on this space's intriguing features and properties. concept uses representation rather than an absolute representation, by comparing set well-trained speakers. advantage approach that instead estimating numerous parameters model speaker, only few number models are estimated. In order optimize performance proposed system, several techniques were assessed possible implementation in various blocks system. As result, best achieved where normalized vector's mutual angle with Minimum normalization method applied conjunction orthogonal virtual case, Equal Error Rate (EER) 0.12% 400 test samples 100 speakers obtained. addition assessment under normal conditions, evaluated abnormal conditions noisy or telephonic speech sequence contamination present. Experiments conducted case demonstrated that, most cases, outperforms systems even shortened sequences. Another major contribution research development more complex capable tackling effectively. other interesting employed. For purpose, novel enrichment construct tackle noise. results experiments part excellent ability conditions. Compared applying led lower error rates all cases low SNR values.