Efficient Integrated Features Based on Pre-trained Models for Speaker Verification

Previous work has explored the application of pre-trained models (PTMs) in speaker verification(SV). Most researchers directly replaced handcrafted features with the universal representations of the PTMs, and jointly fine-tuned PTMs with the downstream SV networks, which undoubtedly discarded important spectral information contained in handcrafted features and also increased the training cost. In this paper, we proposed an efficient feature integration method that utilized a Fine-grained Fusion Module to fuse the multi-layer representations of the PTMs adaptively. Then we integrated the fused representations with handcrafted features to obtain the integrated features, which were subsequently fed into the SV network. The experimental results demonstrated that using the integrated features effectively enhanced the performance of the SV systems, and yielded decent results with no need to fine-tune the PTMs. Moreover, employing full-parameter fine-tuning led to the current optimal results.