This is an implementation of Deep Segment Attentive Embedding on Python 3, Pytorch. The model learn the unified speaker embeddings for utterances of variable duration. It is the 'nlpr_fight' teamwork of Tongdun speaker verification competition.

Python 3.5, Pytorch 0.4.0.

You can download AISHELL to run the code.

You can run sh run.sh for AISHELL, but it's recommended that you run the commands one by one. You need set aishell_data='Your AISHELL Dataset Dir' such as /mnt/DATA/AiShell/data_aishell/wav/ in run.sh.

## make dataset
featdir=fbank/
for x in train dev test; do
    python3 local/aishell_data_prep.py  $aishell_data/$x data/$x $featdir/$x || exit 1;
done 
## make test_pair for verification
for x in dev test; do
    python3 local/make_pairs.py data/$x 1500 1500 || exit 1;
done 

You need build train and dev directory. Each has feats.scp and utt2spk. Each line of feats.scp is "utt_id feats_path" and each line of utt2spk is "utt_id spk_id". The dev directory also need pair.txt. Each line of pair.txt is "utt_id0 utt_id1 feats_path0 feats_path1 label". You can run python3 local/make_pairs.py dev 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 our proposed Deep Segment Attentive Embedding mehod[DSAE]. The model is LSTM or CNN. We also provide different attention strategies.

sh run.sh --stage 2 --seq_training false --model_type lstm --train_type  base_attention | last_state | base_attention ```   
## CNN does not use attention but avgpool.
sh run.sh --stage 2 --seq_training false --model_type cnn 

sh run.sh --stage 2 --seq_training true --model_type lstm | cnn  --train_type base_attention | last_state | multi_attention 
--segment_type  none | all | average