change-detection

The MIT License (MIT)

Copyright (c) 2017 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.

AUTHOR
Ruiqing Yin

Speaker change detection with pyannote.audio

In this tutorial, you will learn how to train and test a speaker change detection model based on MFCCs and LSTMs, using pyannote-change-detection command line tool.

Citation

If you use pyannote-audio for speaker (or audio) change detection, please cite the following paper:

@inproceedings{Yin2017,
  Author = {Ruiqing Yin and Herv\'e Bredin and Claude Barras},
  Title = {{Speaker Change Detection in Broadcast TV using Bidirectional Long Short-Term Memory Networks}},
  Booktitle = {{Interspeech 2017, 18th Annual Conference of the International Speech Communication Association}},
  Year = {2017},
  Month = {August},
  Address = {Stockholm, Sweden},
  Url = {https://github.com/yinruiqing/change_detection}
}

Installation

$ conda create --name py35-pyannote-audio python=3.5 anaconda
$ ource activate py35-pyannote-audio
$ conda install gcc
$ conda install -c conda-forge yaafe
$ pip install "pyannote.audio==0.2.1"
$ pip install pyannote.db.etape

Experimental setup

ETAPE database

This tutorial relies on the ETAPE database. We first need to tell pyannote where the audio files are located:

$ cat ~/.pyannote/db.yml
Etape: /path/to/Etape/corpus/{uri}.wav

Configuration

To ensure reproducibility, pyannote-change-detection relies on a configuration file defining the experimental setup:

$ cat tutorials/change-detection/config.yml
feature_extraction:
   name: YaafeMFCC
   params:
      e: False                   # this experiments relies
      De: True                   # on 11 MFCC coefficients
      DDe: True                  # with 1st and 2nd derivatives
      D: True                    # without energy, but with
      DD: True                   # energy derivatives
      stack: 1

architecture:
   name: StackedLSTM
   params:                       # this experiments relies
     n_classes: 1                # on one LSTM layer (16 outputs)
     lstm: [16]                  # and one dense layer.
     mlp: [16]                   # LSTM is bidirectional
     bidirectional: 'concat'
     final_activation: 'sigmoid'

sequences:
   duration: 3.2                 # this experiments relies on sliding windows
   step: 0.8                     # of 3.2s with a step of 0.8s
   balance: 0.05                 # and balancing neighborhood size of 0.05s

Training

The following command will train the network using the training set of the TV protocol of the ETAPE database. This may take a long time...

$ export EXPERIMENT_DIR=tutorials/change-detection
$ pyannote-change-detection train \       #  
          ${EXPERIMENT_DIR} \             # <experiment_dir>
          Etape.SpeakerDiarization.TV     # <database.task.protocol>
Epoch 1/100
62464/62464 [==============================] - 171s - loss: 0.1543 - acc: 0.9669   
Epoch 2/100
62464/62464 [==============================] - 117s - loss: 0.1375 - acc: 0.9692     
Epoch 3/100
62464/62464 [==============================] - 115s - loss: 0.1376 - acc: 0.9691     
...
Epoch 50/100
62464/62464 [==============================] - 112s - loss: 0.0903 - acc: 0.9724  
...

This will create a bunch of files in TRAIN_DIR (defined below), including plots showing the accuracy epoch after epoch.

Evaluation

Now that the network is trained, we get different models for different epochs. We can evaluate a specific model:

$ export TRAIN_DIR=${EXPERIMENT_DIR}/train/Etape.SpeakerDiarization.TV.train
$ pyannote-change-detection evaluate \
         --epoch=49 \
         ${TRAIN_DIR} \               # <train_dir>
         Etape.SpeakerDiarization.TV  # <database.task.protocol>

This is the expected output:

threshold purity  coverage
--------- ------- --------
0         95.720% 36.603%
0.0526316 95.526% 49.018%
0.105263  95.213% 57.660%
...
0.526316  92.396% 83.756%
...
0.894737  88.155% 91.139%
0.947368  87.468% 92.046%
1         86.929% 92.672%

Testing

We can choose the best model according to the evaluation results and apply it on the development set of the TV protocol of the ETAPE database:

$ pyannote-speech-detection apply \
          --epoch=49 \
          --threshold=0.1 \
          ${TRAIN_DIR} \               #<train_dir>
          Etape.SpeakerDiarization.TV  # <database.task.protocol>

this will create a list of files in APPLY_DIR (defined below) containing segmentation results.

$ export APPLY_DIR=${TRAIN_DIR}/segment/Etape.SpeakerDiarization.TV.development/0.1/
$ head -n 5 $APPLY_DIR/BFMTV_BFMStory_2011-03-17_175900.0.seg
            file_id              seg_id  channel start length
-------------------------------- ------- ------- ----- ------
BFMTV_BFMStory_2011-03-17_175900    0      1      -1     560
BFMTV_BFMStory_2011-03-17_175900    1      1      559    356
BFMTV_BFMStory_2011-03-17_175900    2      1      915    204
BFMTV_BFMStory_2011-03-17_175900    3      1      1119   160
BFMTV_BFMStory_2011-03-17_175900    4      1      1279   398

Going further...

$ pyannote-change-detection --help