음성학적 과학수사 용도의 달팽이관 모사 스펙트럼을 이용하는 딥러닝에 기반한 대화자 식별 알고리즘

초록

In the field of forensic investigation based on limited evidence, it is essential to develop an algorithm that can derive the best performance on a given minimal data. In this study, the authors developed a deep learning algorithm capable of speaker recognition using a short speech of an average of 0.556 seconds. The cochleagram used in the development preserves the existing waveform of the speech signal as much as possible while simulating the preprocessing process in humans auditory system. Therefore, the network can focus on classification task while learning. As a result, when the utterance of one of 71 speakers was input to the 2D convolutional neural network, the average speaker identification accuracy of 10 fold cross-validation was 96.2%. The collected database can be used for later lie detection research because it reflecting the stimulus test used in the lie detection investigation. And expect the developed algorithm can be used to identify new suspects using the ex-convict database.

키워드