train

The MIT License (MIT)

Copyright (c) 2019 CNRS

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

AUTHORS Hervé Bredin - http://herve.niderb.fr

Training on your own dataset with pyannote.audio

In this tutorial, you will learn how to setup your own dataset so that models can be trained on it. We will call this dataset YourDataset in the rest of this tutorial.

Audio files

Though pyannote.audio supports other file formats (it is based on SoundFile), let us assume that YourDataset contains 3 WAV files:

/path/to/your/dataset/audio/file1.wav
/path/to/your/dataset/audio/file2.wav
/path/to/your/dataset/audio/file3.wav

Reference files

Your dataset needs to come with annotations to be of any use for training, in the form of RTTM and UEM files.

"who speaks when" annotations should be provided using the RTTM file format. Each line in this file must follow the following convention:

SPEAKER {uri} 1 {start} {duration} <NA> <NA> {identifier} <NA> <NA>

where {uri} stands for "unique resource identifier" (think of it as the filename), {start} is the start time (elapsed time since the beginning of the file, in seconds) of the speech turn, {duration} is its duration (in seconds) and {identifier} is the unique speaker identifier.

Here what it would look like for YourDataset:

$ cat /path/to/your/dataset/train.rttm
SPEAKER file1 1 0.130 3.880 <NA> <NA> alice <NA> <NA>
SPEAKER file1 1 4.790 0.960 <NA> <NA> alice <NA> <NA>
SPEAKER file1 1 6.190 0.910 <NA> <NA> bob <NA> <NA>
SPEAKER file1 1 7.670 2.340 <NA> <NA> alice <NA> <NA>
SPEAKER file1 1 10.830 2.400 <NA> <NA> carol <NA> <NA>
SPEAKER file1 1 13.670 3.430 <NA> <NA> carol <NA> <NA>
SPEAKER file2 1 17.900 2.210 <NA> <NA> john <NA> <NA>
SPEAKER file2 1 20.370 0.760 <NA> <NA> jack <NA> <NA>
SPEAKER file2 1 21.560 3.410 <NA> <NA> john <NA> <NA>
SPEAKER file3 1 25.500 3.410 <NA> <NA> hugh <NA> <NA>

It is possible that only parts of your files are annotated. This is the role of the UEM file: telling pyannote-audio which part were actually annotated.

If you do not provide this file, pyannote-audio assumes that the whole file was annotated and therefore everything that is outside of a speech turn is considered non-speech.

If you do provide this file, pyannote-audio will only consider as non-speech those regions that are within the limits defined in the UEM file.

Each line in this file must follow the following convention:

{uri} 1 {start} {end}

Here is what it might look like for YourDataset

$ cat /path/to/your/dataset/train.uem
file1 1 0.000 120.0
file1 1 130.0 240.0
file2 1 0.000 300.0
file3 1 60.0 300.0

Configuration file

Once everything is ready, you can update (or create if it does not exist) file /path/to/database.yml like this:

Databases:
  YourDataset: /path/to/your/dataset/audio/{uri}.wav

Protocols:
  YourDataset:
    SpeakerDiarization:
      YourProtocol:
        train:
          annotation: /path/to/your/dataset/train.rttm
          annotated: /path/to/your/dataset/train.uem

... and tell pyannote.database about this file:

$ export PYANNOTE_DATABASE_CONFIG=/path/to/database.yml

Congratulations: you have just defined a new pyannote.database protocol (called YourDataset.SpeakerDiarization.YourProtocol) that can be used in pyannote.audio.

All you have to do now is to replace AMI.SpeakerDiarization.MixHeadset by YourDataset.SpeakerDiarization.YourProtocol in the tutorials explaining how to train models...

Note that you should probably add a development set for validating the models (and optionally a test set for proper evaluation):

Databases:
  YourDataset: /path/to/your/dataset/audio/{uri}.wav

Protocols:
  YourDataset:
    SpeakerDiarization:
      YourProtocol:
        train:
          annotation: /path/to/your/dataset/train.rttm
          annotated: /path/to/your/dataset/train.uem
        development:
          annotation: /path/to/your/dataset/development.rttm
          annotated: /path/to/your/dataset/development.uem
        test:
          annotation: /path/to/your/dataset/test.rttm
          annotated: /path/to/your/dataset/test.uem

That's all folks!