VBx-training-recipe

Training recipe for VBx repository. For results and more info please see VBx.

Installation

This is Kaldi based recipe, using most of the tools from egs/sre16/v2, therefore it is required to have Kaldi compiled.

From Kaldi root dir:

cd egs/sre16
git clone https://github.com/phonexiaresearch/VBx-training-recipe.git
cd VBx-training-recipe

For NN training it is also needed to compile binary which converts egs into arks - utils/nnet3-copy-egs-to-feats.cc. Please, copy this source code into KALDI/src/nnet3bin and compile it.

cp utils/nnet3-copy-egs-to-feats.cc ../../../src/nnet3bin/

Please see run.sh script for basic overview, minimal modification is to set correct VoxCeleb directories for your setup here.

When set, you can run the main script:

./run.sh

For more details see run.sh. If you are familiar with sre16/v2 recipe, it should be straightforward.

Citations

In case of using the software please cite:

F. Landini, J. Profant, M. Diez, L. Burget: Bayesian HMM clustering of x-vector sequences (VBx) in speaker diarization: theory, implementation and analysis on standard tasks

Contact

If you have questions regarding the training recipe, email [email protected] or [email protected].