wenwanchen/kaldi-ivector

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

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.