Trudności związane ze słuchowym i automatycznym rozpoznawaniem osób na podstawie głosu: analiza przypadków głosów bardzo podobnych
Challenges for Auditory and Automatic Speaker Recognition: Evaluating Cases of Highly Similar Voices

Summary/Abstract: Identical twins present a difficult case for both auditory and machine speaker recognition. This paper addresses this challenge and presents the findings of two studies: an auditory speaker discrimination test and a machine-based task using forensic automatic speaker recognition (ASR) system. The outcomes of the perceptual judgement task were compared with the log-likelihood ratios (LLRs) yielded by an x-vector-based speaker recognition system. Although the task was given to lay listeners, as opposed to forensic phonetic experts, the results appear to be congruent with the scores yielded by a state-of-the-art automatic system. The human raters were more accurate in judging same-speaker pairs than different-speaker pairs. The machine approach showed better performance in both conditions tested as compared to human listeners. Overall, the voices that were difficult for human listeners were different from those that the ASR system struggled with.

• Details
• Contents