520liuxing/dover-lap

Python package for combining diarization system outputs.

DOVER-Lap

Official implementation for DOVER-Lap: A method for combining overlap-aware diarization outputs.

Installation

pip install dover-lap

How to run

After installation, run

dover-lap [OPTIONS] OUTPUT_RTTM [INPUT_RTTMS]...

Example:

dover-lap egs/ami/rttm_dl_test egs/ami/rttm_test_*

Usage instructions

Usage: dover-lap [OPTIONS] OUTPUT_RTTM [INPUT_RTTMS]...

  Apply the DOVER-Lap algorithm on the input RTTM files.

Options:
  --custom-weight TEXT            Weights for input RTTMs
  --dover-weight FLOAT            DOVER weighting factor  [default: 0.1]
  --weight-type [rank|custom]     Specify whether to use rank weighting or
                                  provide custom weights  [default: rank]

  --tie-breaking [uniform|all]    Specify whether to assign tied regions to
                                  all speakers or divide uniformly  [default:
                                  all]

  --second-maximal                If this flag is set, run a second iteration
                                  of the maximal matching for greedy label
                                  mapping  [default: False]

  --sort-first                    If this flag is set, sort inputs by DER
                                  first before label mapping (only applicable
                                  when label mapping type is hungarian)
                                  [default: False]

  --label-mapping [hungarian|greedy]
                                  Choose label mapping algorithm to use
                                  [default: greedy]

  --random-seed INTEGER
  -c, --channel INTEGER           Use this value for output channel IDs
                                  [default: 1]

  -u, --uem-file PATH             UEM file path
  --help                          Show this message and exit.

Note:

1. If --weight-type custom is used, then --custom-weight must be provided. For example:

dover-lap egs/ami/rttm_dl_test egs/ami/rttm_test_* --weight-type custom --custom-weight '[0.4,0.3,0.3]'

2. label-mapping can be set to greedy (default) or hungarian, which was the mapping technique originally proposed in DOVER.

Results

We provide a sample result on the AMI mix-headset test set. The results can be obtained using spyder, which is automatically installed with dover-lap:

dover-lap egs/ami/rttm_dl_test egs/ami/rttm_test_*
spyder egs/ami/ref_rttm_test egs/ami/rttm_dl_test

and similarly for the input hypothesis. The DER results are shown below.

	MS	FA	Conf.	DER
Overlap-aware VB resegmentation	9.84	2.06	9.60	21.50
Overlap-aware spectral clustering	11.48	2.27	9.81	23.56
Region Proposal Network	9.49	7.68	8.25	25.43
DOVER-Lap (Hungarian mapping)	9.81	2.76	8.17	20.73
DOVER-Lap (Greedy mapping)	9.71	3.02	7.68	20.40

Running time

The algorithm is implemented in pure Python with NumPy for tensor computations. The time complexity is expected to increase exponentially with the number of inputs, but it should be reasonable for combining up to 10 input hypotheses.

For smaller number of inputs (up to 5), the algorithm should take only a few seconds to run on a laptop.

Combining 2 systems with DOVER-Lap

DOVER-Lap is meant to be used to combine more than 2 systems, since black-box voting between 2 systems does not make much sense. Still, if 2 systems are provided as input, we fall back on the Hungarian algorithm for label mapping, since it is provably optimal for this case. Both the systems are assigned equal weights, and in case of voting conflicts, the region is equally divided among the two labels. This is not the intended use case and will almost certainly lead to performance degradation.

Citation

@article{Raj2021Doverlap,
  title={{DOVER-Lap}: A Method for Combining Overlap-aware Diarization Outputs},
  author={D.Raj and P.Garcia and Z.Huang and S.Watanabe and D.Povey and A.Stolcke and S.Khudanpur},
  journal={2021 IEEE Spoken Language Technology Workshop (SLT)},
  year={2021}
}

Contact

For issues/bug reports, please raise an Issue in this repository, or reach out to me at [email protected].