Skip to content

Latest commit

 

History

History
executable file
·
75 lines (47 loc) · 2.35 KB

readme.md

File metadata and controls

executable file
·
75 lines (47 loc) · 2.35 KB
Raw

Speaker Recognition Demo

This code is modified based on https://github.com/clovaai/voxceleb_trainer , include the pretrain model. This is just a simply version of that repository for the 'demo' purpose.

What can it do ?

This code can output the recognition scores between the test utterance (the utterance you want to know the the speaker id) and all enrollment utterances (the utterances you have registed and already known the speaker id) based on the person's unique voiceprint.

How to use ?

Clone the code:

git clone https://github.com/TaoRuijie/SpeakerRecognitionDemo

Input:

Put the wav files you have registed in the data/enrollment_audio folder. Put the wav file you want to test in data/test_audio folder (You can put multiple test audio file, each test audio will generate one score list).

Command:

python demoSpeakerNet.py

Output:

The score is the speaker recognition score between the test utterance and all enrollment utterances, the score is higher, the utterance tends to come from the same speaker (The highest is 0.0).

For instance: There are 7 audios in data/enrollment_audio and 1 audio in data/test_audio. You can get the following output:

A1
A2 -0.8122
D2 -1.2551
D1 -1.2641
B1 -1.3041
C2 -1.3062
B2 -1.3246
C1 -1.3756

A1 and A2 come from the same speaker, so they have the highest score.

Other issue:

  • Audio file format

    Current only wav, any sampling rate, single/dual channel is fine.

  • Requirements

    Pytorch (>=1.6.0), pandas, soundfile package

  • GPU

    This is just the demo code, so the default setting is to use cpu only. It will not be very slow.

  • Threshold setting

    It depends on the dataset. Just for suggestion: usually the score between 0 to -1.0 can be viewed as the same speaker, score small than -1.0 can be viewed as different speaker.

  • Reference

    Please check their paper for more details:

     @inproceedings{chung2020in,
   title={In defence of metric learning for speaker recognition},
   author={Chung, Joon Son and Huh, Jaesung and Mun, Seongkyu and Lee, Minjae and Heo, Hee Soo and Choe, Soyeon and Ham, Chiheon and Jung, Sunghwan and Lee, Bong-Jin and Han, Icksang},
   booktitle={Interspeech},
   year={2020}
 }