This is an implementation of Speaker Verification on Python 3, Pytorch. The model can be trained by the ge2e loss or classification loss.

Python 3.5, Pytorch 1.0.0.

You can download AISHELL to run the code.

You need build train, dev and test directory. Each has feats.scp and utt2spk. The test directory need pair.txt. Each line of pair.txt is "utt_id0 feats_path0 utt_id1 feats_path1 label". You can run python3 data/make_pairs.py test 1500 1500 to randomly build the pair.txt.

We provide two training methods. One is based on "Generalized End-to-End Loss for Speaker Verfication"GE2E. Anonther is x-vector(classification loss). The model is LSTM or CNN. We also provide different attention strategies and different margin strategies.

sh run.sh --loss_type class_softmax | ge2e_softmax --model_type lstm | cnn --att_type  base_attention | last_state | multi_attention --margin_type Softmax ```