vvestman / pytorch-ivectors

GPU accelerated implementation of i-vector extractor training using PyTorch. Requires Kaldi for feature extraction and UBM training. An example script is provided for VoxCeleb data.

GPU accelerated PyTorch implementation of frame posterior computation and i-vector extractor training.

Kaldi is required for MFCC extraction and UBM training.

Steps to run example script with VoxCeleb data:

• Move kaldi/egs/voxceleb/v1/extract_feats_and_train_ubm.sh to the corresponding folder in your Kaldi installation
• In extract_feats_and_train_ubm.sh, update output_dir, voxceleb1_root, and voxceleb2_root.
  • If you are using newer version of VoxCeleb1 (1.1), you might have to modify kaldi/egs/voxceleb/v1/local/make_voxceleb1.pl as the data organization is different than in the original VoxCeleb release.
• run extract_feats_and_train_ubm.sh
• update DATA_FOLDER in run_voxceleb_ivector.py
• install and activate compatible conda environment
  • environment.yml has all the needed packages
  • Main requirements: Python (>3.6), PyTorch(>1.1), NumPy, SciPy, PyKaldi
• run run_voxceleb_ivector.py

For more details: http://dx.doi.org/10.21437/Interspeech.2019-1955

@inproceedings{Vestman2019,
  author={Ville Vestman and Kong Aik Lee and Tomi H. Kinnunen and Takafumi Koshinaka},
  title={{Unleashing the Unused Potential of i-Vectors Enabled by GPU Acceleration}},
  year=2019,
  booktitle={Proc. Interspeech 2019},
  pages={351--355},
  doi={10.21437/Interspeech.2019-1955},
  url={http://dx.doi.org/10.21437/Interspeech.2019-1955}
}