https://github.com/swshon/voxceleb-ivector

Voxceleb1 i-vector based speaker recognition system
https://github.com/swshon/voxceleb-ivector

i-vector kaldi speaker-embedding speaker-identification speaker-recognition speaker-verification voxceleb voxceleb1

Last synced: 26 days ago
JSON representation

Voxceleb1 i-vector based speaker recognition system

• Host: GitHub
• URL: https://github.com/swshon/voxceleb-ivector
• Owner: swshon
• License: mit
• Created: 2018-04-09T19:14:03.000Z (about 6 years ago)
• Default Branch: master
• Last Pushed: 2018-05-22T21:54:54.000Z (about 6 years ago)
• Last Synced: 2024-02-21T05:36:00.948Z (4 months ago)
• Topics: i-vector, kaldi, speaker-embedding, speaker-identification, speaker-recognition, speaker-verification, voxceleb, voxceleb1
• Language: Perl
• Size: 2.16 MB
• Stars: 41
• Watchers: 2
• Forks: 11
• Open Issues: 1
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Lists

• awesome-diarization - voxceleb-ivector - ivector?style=social) | i-vector |Perl | Voxceleb1 i-vector based speaker recognition system. | (Software / Speaker embedding)
• awesome-asv-antispoofing - voxceleb-ivector - ivector?style=social) | i-vector |Perl | Voxceleb1 i-vector based speaker recognition system. | (Software / Speaker embedding)

README

        

# Speaker Verification task in Voxceleb1 dataset

This repository contains simple scripts for a training i-vector speaker recognition system on Voxceleb1[1] dataset using Kaldi. It was modified based on run.sh file on Kaldi/egs/sre10.

# Requirement

* Kaldi Toolkit

# How to use

1. Move all files to {kaldi_root}/egs/sre10 folder

2. Modify dataset directories and parameters in run.sh file to fit in your machine.

3. Run run.sh file

# Result

The 2048 component GMM-UBM and 600-dimensional i-vector extractor were trained using voxceleb1 training data for verification task. Training parameter is almost same compared to sre10 baseline on Kaldi egs.

GMM-2048 CDS eer : 15.39%


GMM-2048 LDA+CDS eer : 8.103%


GMM-2048 PLDA eer : 5.446%


# Note

The Voxceleb1 dataset, a large-scale speaker identification dataset was published in 2017 with speaker embedding baseline[1] and reported i-vector shows 8.8% EER. The i-vector was extracted using 1024 component GMM-UBM, so the EER is fairly worse compared to the result above.

# Reference

[1] A. Nagraniy, J. S. Chung, and A. Zisserman, “VoxCeleb: A large-scale speaker identification dataset,” in Interspeech, 2017, pp. 2616–2620.

* CSV file in data folder created from here 

(https://github.com/pyannote/pyannote-db-voxceleb/blob/master/scripts/prepare_data.ipynb)