Crossed-Time Delay Neural Network for Speaker Recognition

TL;DR

The paper introduces Crossed-Time Delay Neural Network (CTDNN), a novel architecture that enhances speaker recognition performance by employing multiple context-sized delay units, outperforming existing TDNN models in verification and identification tasks.

Contribution

The paper proposes a new CTDNN structure inspired by CNN multi-filters, significantly improving speaker recognition accuracy and training efficiency over traditional TDNN and FTDNN models.

Findings

Outperforms TDNN with 2.6% EER reduction on VoxCeleb1.

Achieves 90.4% identification accuracy in few-shot scenarios.

Demonstrates 36% accuracy improvement over FTDNN.

Abstract

Time Delay Neural Network (TDNN) is a well-performing structure for DNN-based speaker recognition systems. In this paper we introduce a novel structure Crossed-Time Delay Neural Network (CTDNN) to enhance the performance of current TDNN. Inspired by the multi-filters setting of convolution layer from convolution neural network, we set multiple time delay units each with different context size at the bottom layer and construct a multilayer parallel network. The proposed CTDNN gives significant improvements over original TDNN on both speaker verification and identification tasks. It outperforms in VoxCeleb1 dataset in verification experiment with a 2.6% absolute Equal Error Rate improvement. In few shots condition CTDNN reaches 90.4% identification accuracy, which doubles the identification accuracy of original TDNN. We also compare the proposed CTDNN with another new variant of TDNN, FTDNN, which shows that our model has a 36% absolute identification accuracy improvement under few shots condition and can better handle training of a larger batch in a shorter training time, which better utilize the calculation resources. The code of the new model is released at https://github.com/chenllliang/CTDNN