Speaker Recognition System in Forensic Conditions: The Calibration and Evaluation of the Likelihood Ratio
Miranti Indar Mandasari · Radboud Repository (Radboud University) · 2018
This thesis contributes on evaluation and enhancement of a modern speaker recognition system such that it can be more usable in forensics. The speaker recognition system is a system developed at Radboud University Nijmegen (RUN) and based on the-state-of-the-art i-vector framework. Evaluation of the speaker recognition system was carried out thoroughly by taking into account aspects in both recognition and calibration performances. Duration and noise conditions are two forensic-motivated conditions investigated within the evaluation. A new approach in calibration is proposed in order to tackle the two aforementioned variabilities. The approach is called quality measure function (QMF) calibration. Here, quality measurements from speech, i.e., signal to noise ratio (SNR) and duration of active speech, are incorporated to the traditional linear calibration. The QMF calibration was then comprehensively evaluated. Results show that this proposed method results in improvement in recognition and calibration aspects. In addition, this thesis includes transfer techniques and approaches from speaker recognition field, where the concept of calibration is introduced to the face recognition community.