Same Speaker Identification with Deep Learning and Application to Text-Dependent Speaker Verification

Abstract

In human-based visual and auditory forensic speaker recognition or comparison, experts make judgments on whether or not two speech samples are from the same speaker. They compare specific parts of the same phoneme where the speaker's characteristics appear. Inspired by this process, we propose a Same Speaker Identification Deep Neural Network (SSI-DNN) that identifies whether or not two speech samples are uttered by the same speaker by focusing on the same phonemes, which have very short durations. The architecture of the proposed SSI-DNN is based on ResNet, with the input extended to be two speech samples, and the output is the identification result. Because the proposed method does not use speaker embedding or speaker models, it can be trained on a small number of utterances for each speaker. To evaluate the verification performance, we conducted speaker verification experiments for five Japanese single vowels, which are 0.12 s each, and a short phrase “oi,” which means “hey,” in a speech database for the forensic speaker recognition constructed by the National Research Institute of Police Science in Japan. The experimental results showed that when two speech samples were uttered 30 min apart, the equal error rate (EER) for single vowels was 7.4%, 1.2% for the five-vowel score combination, and 4.0% for “oi.” With about a three-month interval, the EER was 12.8% for single vowels, 4.2% for the five-vowel score combination, and 8.1% for “oi.”