Whisper40: A Multi-Person Chinese Whisper Speaker Recognition Dataset Containing Same-Text Neutral Speech

Abstract

Whisper speaker recognition (WSR) has received extensive attention from researchers in recent years, and it plays an important role in medical, judicial, and other fields. Among them, the establishment of a whisper dataset is very important for the study of WSR. However, the existing whisper dataset suffers from the problems of a small number of speakers, short speech duration, and lack of neutral speech with the same-text as the whispered speech in the same dataset. To address this issue, we present Whisper40, a multi-person Chinese WSR dataset containing same-text neutral speech spanning around 655.90 min sourced from volunteers. In addition, we use the current state-of-the-art speaker recognition model to build a WSR baseline system and combine the idea of transfer learning for pre-training the speaker recognition model using neutral speech datasets and transfer the empirical knowledge of specific network layers to the WSR system. The Whisper40 and CHAINs datasets are then used to fine-tune the model with transferred specific layers. The experimental results show that the Whisper40 dataset is practical, and the time delay neural network (TDNN) model performs well in both the same/cross-scene experiments. The equal error rate (EER) of Chinese WSR after transfer learning is reduced by 27.62% in comparison.