Deep Self-Supervised Hierarchical Clustering for Speaker Diarization

This repository is code used in our paper:

Deep Self-Supervised Hierarchical Clustering for Speaker Diarization
Prachi Singh, Sriram Ganapathy
Accepted in Interspeech 2020

Overview

The recipe consists of:

• extracting x-vectors using pretrained model
• Performing self-supervised clustering for diarization on Callhome set
• Evaluating results using Diarization error rate (DER)

Prerequisites

The following packages are required to run the baseline.

• Python >= 3.6
• Kaldi
• Pytorch >= 0.4.1
• dscore

Getting started

• clone the repository:

git clone https://github.com/iiscleap/self_supervised_AHC.git

• Install Kaldi. If you are a Kaldi novice, please consult the following for additional documentation:
  • Kaldi tutorial
  • Kaldi for Dummies tutorial
• Go to cloned repository and copy kaldi path in path.sh given as:

$ local_dir="Full_path_of_cloned_repository"
$ echo "export KALDI_ROOT="/path_of_kaldi_directory/kaldi" >> $local_dir/path.sh
$ echo "export KALDI_ROOT="/path_of_kaldi_directory/kaldi" >> $local_dir/tools_dir/path.sh

• Input x-vectors features are obtained using Kaldi Callhome-diarization. Pre-trained x-vector model and plda model including global mean and PCA transform needed for training are given in tools_diar/callhome_xvector_models:
• Performance is evaluated using dscore. Download all the required dependencies in the same python environment.

Implementation

X-vectors Extraction

• This step is to run kaldi diarization pipeline till x-vector extraction using pre-trained model
• Additionally it will convert x-vectors in ark format into numpy format to run in pytorch. It will also convert kaldi plda model into pickle format.
• Replace "data_root" with path of callhome dataset in tools_diar/run_extract_xvectors.sh
• Run following commands:

$ local_dir="Full_path_of_cloned_repository"
$ cd $local_dir/tools_diar
$ bash run_extract_xvectors.sh

DNN Training

• xvec_ahc_train.py is code for DNN training
• run_xvec_ahc.sh calls DNN training script
• Update training parameters in run_xvec_ahc.sh NOTE that by default Kaldi scripts are configured for execution on a grid using a submission engine such as SGE or Slurm. If you are running the recipes on a single machine, make sure to edit cmd.sh and tools_dir/cmd.sh so that the line

   $ export train_cmd="queue.pl"

reads

   $ export train_cmd="run.pl"

• Execute following commands:

$ local_dir="Full_path_of_cloned_repository"
$ cd $local_dir
$ bash run_xvec_ahc.sh $local_dir  --TYPE parallel --nj <number of jobs> --which_python <python_env_with_all_installed_libraries>

Note: --TYPE parallel (when running multiple jobs simultaneoulsy)

Evaluation

• Diarization Error Rate is used as performance metric
• Scripts in dscore generates filewise DER.
• Go to cloned repo and run following command for evaluation

$ local_dir="Full_path_of_cloned_repository"
$ cd $local_dir
$ find $local_dir/rttm_ground/*.rttm > ref_callhome.scp
$ cd tool_diar/
$ bash gen_rttm.sh --stage <1/2> --modelpath <path of model to evaluate> --which_python <python_env_with_all_installed_libraries>

Note: --stage 1 (using ground truth number of speakers), --stage 2 (using threshold based number of clusters)

Output

• Generates der.scp in modelpath which contains filewise DER and other metric like JER.

Cite

@article{singh2020deep, title={Deep Self-Supervised Hierarchical Clustering for Speaker Diarization}, author={Singh, Prachi and Ganapathy, Sriram}, journal={arXiv preprint arXiv:2008.03960}, year={2020} }

Contact

If you have any comment or question, please contact prachisingh@iisc.ac.in