GE2E loss speaker embedding

Keras-based Implementation

• This is the implementation of Google's GE2E Speaker Verification based on tensorflow keras. It has also intuitive close-voice-web-DEMO. We chose EER as a barometer of our Speaker Embedding System. Using GE2E loss, various categorization could be(like age, country, emotion, etc).

DATA

• vox celeb1, vox celeb2, librispeech (총 9500명 화자)

How to Run

• Change the MODE parameter in config/config.json to "TRAIN", "INFER" or "EXRPOT" before running.

  1. Train

  python ge2e.py

  2. Infer

  --ckpt arg를 "latest" 로 하면 최신 ckpt가 설정됨, 특정 ckpt를 사용하고자 하면 full path 입력

  -실행시 wavdir 에 대해 preprocessing이 먼저 진행됨 (path to wav directory + '_preprocessed' 폴더가 생성됨; 상위 폴더 권한 설정 유의 바람)

  -emb_savedir아래에 ckpt 별 emb폴더가 생성되므로 emb_savedir에 ckpt epoch 기재할 필요 x

  python ge2e.py --wavdir [path to wav directory] --emb_savedir [path to save embeddings] --ckpt [exact path of the ckpt to use]

  3. Export

  python ge2e.py