zhangxincheng / Discriminative-Speaker-Embedding

Learning Discriminative Speaker Embedding by Improving Aggregation Strategy and Loss Function for Speaker Verification

Discriminative Speaker Embedding

This repository contains the code and pre-trained models for our paper 'Learning Discriminative Speaker Embedding by Improving Aggregation Strategy and Loss Function for Speaker Verification'.

Data preparation

Use the data_prepare.sh script to prepare VoxCeleb1 and VoxCeleb2 dataset.

The Fbank features of the training input are made by data_utils/kaldi/egs/Vox/v1/make_Vox_feats.sh

Training

python trainSpeakerNet.py \
--config config/exp_ResNetSE34L_vox1_trainSeg_MSA-NeXtVLAD_AAM+APL_MT234.yaml 

Evaluation

python trainSpeakerNet.py \
--config config/test_ResNetSE34L_vox1_trainSeg_MSA-NeXtVLAD_AAM+APL_MT234.yaml --eval