A Lightweight Speaker Verification Model For Edge Device
Tingwei Chen, Chia-Ping Chen, Chung-Li Lu, Bo-Cheng Chan, Yu-Han Cheng, Hsiang-Feng Chuang, Wei-Yu Chen · 2023
In this paper, we investigate lightweight models for automatic speaker verification (ASV) that can be feasibly implemented on edge devices. The proposed model architecture includes an improved and re-parameterized VGG-based front-end stem (RepVGG) and an ECAPA-TDNNLite-like backbone. We refer the proposed model by IM ECAPA Rep-TDNNLite. We further apply the decoupled knowledge distillation approach to improve the performance of the lightweight model. On the open VoxCeleb1-O test set, the proposed IM ECAPA Rep-TDNNLite achieves 1.76% equal error rate (EER) and 0.13 minimum decision cost function (minDCF) with 1.31M model parameters and 1.45G FLOPs. The performance can be improved to 1.66% EER and 0.12 minDCF by a different RepVGG front-end with 1.46M parameters and 2.67G FLOPs. Thus, the proposed light-weight model is capable of providing good performance on edge devices with limited computation resources.