ge2eloss-svf

Tensorflow implementation of Generalized End-to-End Loss for speaker verification, proposed by google on 2017 in https://arxiv.org/pdf/1710.10467.pdf

Environment

python3.7,tensorflow-gpu==1.14.0

Dataset

We use our own dataset which belong to the game customer service.The audio sampled with 8k.We collected the audios with 30000 speakers and segement the audio to short but greater than 6s with Vad technology.

To preprocess the data with dataset/generate_meta.py

Training script

python train.py \
--train_meta_files \
/tmp/meta.txt \
--epochs 10000 \
--max_keep_model 100 \
--batch_size 100 \
--utterances_per_speaker 10 \
--lr_decay_type constant \
--process_name sf_wj \
--gpu_devices 0,1,2,3,4 \
--thread_num 10 \
--buffer_size 10 \
--log_dir /tmp/svf/log_wj \
--checkpoint /tmp/svf/checkpoint_wj

Training result

Calculate EER

1.eval_test/v2/prepare_eval_data/test_data_utterance_vector.py
2.eval_test/v2/eval_eer.py

Test result

dataset-audio>=5s	test-data with 240 speakers not found in train-data.Each speaker must have at least 10 pieces of 5s + audio.Sample rate is 8k	testdata from aishell.Resample to 8k	traindata in aishell.Resample to 8k	magicdata_mandarin_chinese's test data resample to 8k
eer	0.018	0.04	0.025	0.04

Reference

https://github.com/CorentinJ/Real-Time-Voice-Cloning/tree/master/encoder