Repository files navigation

EEND (End-to-End Neural Diarization) is a neural-network-based speaker diarization method.

• BLSTM EEND (INTERSPEECH 2019)
  • https://www.isca-speech.org/archive/Interspeech_2019/abstracts/2899.html
• Self-attentive EEND (ASRU 2019)
  • https://ieeexplore.ieee.org/abstract/document/9003959/

The EEND extension for various number of speakers is also provided in this repository.

• Self-attentive EEND with encoder-decoder based attractors
  • https://arxiv.org/abs/2005.09921

• NVIDIA CUDA GPU
• CUDA Toolkit (8.0 <= version <= 10.1)

cd toolsmake

• This command builds kaldi attools/kaldi
  • if you want to use pre-build kaldi

    cd toolsmake KALDI=<existing_kaldi_root>

    This option make a symlink attools/kaldi
• This command extracts miniconda3 attools/miniconda3, and creates conda envirionment named 'eend'
• Then, installs Chainer and cupy into 'eend' environment
  • use CUDA in/usr/local/cuda/
    • if you need to specify your CUDA path

      cd toolsmake CUDA_PATH=/your/path/to/cuda-8.0

      This command installs cupy-cudaXX according to your CUDA version.Seehttps://docs-cupy.chainer.org/en/stable/install.html#install-cupy

• Modifyegs/mini_librispeech/v1/cmd.sh according to your job schedular.If you use your local machine, use "run.pl".If you use Grid Engine, use "queue.pl"If you use SLURM, use "slurm.pl".For more information about cmd.sh seehttp://kaldi-asr.org/doc/queue.html.

cd egs/mini_librispeech/v1./run_prepare_shared.sh

./run.sh

• If you use encoder-decoder based attractors [3], modifyrun.sh to useconfig/eda/{train,infer}.yaml
• SeeRESULT.md and compare with your result.

• Modifyegs/callhome/v1/cmd.sh according to your job schedular.If you use your local machine, use "run.pl".If you use Grid Engine, use "queue.pl"If you use SLURM, use "slurm.pl".For more information about cmd.sh seehttp://kaldi-asr.org/doc/queue.html.
• Modifyegs/callhome/v1/run_prepare_shared.sh according to storage paths of your corpora.

cd egs/callhome/v1./run_prepare_shared.sh# If you want to conduct 1-4 speaker experiments, run below.# You also have to set paths to your corpora properly../run_prepare_shared_eda.sh

./run.sh

local/run_blstm.sh

./run_eda.sh

[1] Yusuke Fujita, Naoyuki Kanda, Shota Horiguchi, Kenji Nagamatsu, Shinji Watanabe, "End-to-End Neural Speaker Diarization with Permutation-free Objectives," Proc. Interspeech, pp. 4300-4304, 2019

[2] Yusuke Fujita, Naoyuki Kanda, Shota Horiguchi, Yawen Xue, Kenji Nagamatsu, Shinji Watanabe, "End-to-End Neural Speaker Diarization with Self-attention," Proc. ASRU, pp. 296-303, 2019

[3] Shota Horiguchi, Yusuke Fujita, Shinji Watanabe, Yawen Xue, Kenji Nagamatsu, "End-to-End Speaker Diarization for an Unknown Number of Speakers with Encoder-Decoder Based Attractors," Proc. INTERSPEECH, 2020

@inproceedings{Fujita2019Interspeech, author={Yusuke Fujita and Naoyuki Kanda and Shota Horiguchi and Kenji Nagamatsu and Shinji Watanabe}, title={{End-to-End Neural Speaker Diarization with Permutation-free Objectives}}, booktitle={Interspeech}, pages={4300--4304} year=2019}