Speaker verification over the telephone

摘要: Speaker verification has been the subject of active research for many years, yet despite these eAorts and promising results on laboratory data, speaker performance over telephone remains below that required applications. This experimental study aimed to quantify recognition out context any specific application, as a function factors more-or-less acknowledged aAect accuracy. Some issues addressed are: model (Gaussian mixture models are compared with phone-based models), influence amount content training test data performance; degradation due aging how can this be counteracted by using adaptation techniques; achievable levels text-dependent text-independent modes. These other were large corpus read spontaneous speech (over 250 hours collected from 100 target speakers 1000 imposters) in French, designed recorded purpose study. On lowest equal error rate is 1% mode when two trials allowed per attempt minimum 1.5 s trial. ” 2000 Elsevier Science B.V. All rights reserved.