https://github.com/idiap/kaldi-ivector

Extension to Kaldi implementing the standard i-vector hyperparameter estimation and i-vector extraction procedure
https://github.com/idiap/kaldi-ivector

Last synced: 26 days ago
JSON representation

Extension to Kaldi implementing the standard i-vector hyperparameter estimation and i-vector extraction procedure

• Host: GitHub
• URL: https://github.com/idiap/kaldi-ivector
• Owner: idiap
• License: other
• Created: 2016-01-07T08:33:37.000Z (over 8 years ago)
• Default Branch: master
• Last Pushed: 2018-02-23T06:46:53.000Z (over 6 years ago)
• Last Synced: 2024-02-21T05:35:42.596Z (4 months ago)
• Language: C++
• Size: 57.6 KB
• Stars: 88
• Watchers: 12
• Forks: 25
• Open Issues: 3
• Metadata Files:
  • Readme: README.md
  • License: COPYING

Lists

• awesome-diarization - kaldi-ivector - ivector?style=social) | i-vector | C++ & Perl | Extension to Kaldi implementing the standard i-vector hyperparameter estimation and i-vector extraction procedure. | (Software / Speaker embedding)
• awesome-asv-antispoofing - kaldi-ivector - ivector?style=social) | i-vector | C++ & Perl | Extension to Kaldi implementing the standard i-vector hyperparameter estimation and i-vector extraction procedure. | (Software / Speaker embedding)

README

        
# UPDATED: 20/02/2018

This is the README to the Idiap's implementation of the i-vector

system for Kaldi. It contains information about the package, implementation

details, installation and compilation.

## General information

This implementation of the i-vector system is based on the

standard i-vector extraction procedure. It contains code to estimate

the T-matrix with the conventional EM algorithm for estimation of

Eigenvoice matrices, estimate i-vectors given the T-matrix, features

and corresponding posteriors. 

## Data structures

The classes for T-matrix and sufficient statistics are modifications

to the classes already present in Kaldi. Some irrelevant members are

removed.

The i-vector is still a kaldi::Vector and is compatible with the 

LDA and PLDA backends already available in Kaldi.

## Compilation

To compile the package simply follow the 2 steps

1. export the path to kaldi souce in the environment variable $KALDI_DIR

```

export KALDI_DIR=/home/username/kaldi-trunk/

```

2. Run make in the src/ directory

```

cd src/

make

```

Now, the binaries should have been created in the src/ivectorbin/

folder.

## Kaldi recipe

The recipe equivalent to the kaldi recipe to train and

test a speaker recognition system for NIST SRE 2008 dataset

is available in the scripts folder. The file scripts/run.sh

is the main recipe that calls other scripts from within. 

## References

The implementation is based on the i-vector systems in 

[1] Glembek, Ondřej, et al. "Simplification and optimization of i-vector extraction." Acoustics, Speech and Signal Processing (ICASSP), 2011 IEEE International Conference on. IEEE, 2011.

[2] Madikeri, Srikanth. "A hybrid factor analysis and probabilistic pca-based system for dictionary learning and encoding for robust speaker recognition." Odyssey Workshop. 2012.