The MIT License (MIT)
Copyright (c) 2017-2019 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 Ruiqing Yin Hervé Bredin - http://herve.niderb.fr
In this tutorial, you will learn how to train, validate, and apply a speaker change detector based on MFCCs and LSTMs, using pyannote-change-detection
command line tool.
If you use pyannote-audio
for speaker 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}
}
$ source activate pyannote
$ pip install pyannote.db.odessa.ami
$ pip install pyannote.db.musan
This tutorial relies on the AMI and MUSAN databases. We first need to tell pyannote
where the audio files are located:
$ cat ~/.pyannote/database.yml
Databases:
AMI: /path/to/ami/amicorpus/*/audio/{uri}.wav
MUSAN: /path/to/musan/{uri}.wav
Have a look at pyannote.database
documentation to learn how to use other datasets.
To ensure reproducibility, pyannote-change-detection
relies on a configuration file defining the experimental setup:
$ cat tutorials/models/speaker_change_detection/config.yml
task:
name: SpeakerChangeDetection
params:
duration: 2.0 # sequences are 2s long
collar: 0.100 # upsampling collar = 100ms
non_speech: False # do not try to detect non-speech/speaker changes
batch_size: 64 # 64 sequences per batch
per_epoch: 1 # one epoch = 1 day of audio
data_augmentation:
name: AddNoise # add noise on-the-fly
params:
snr_min: 10 # using random signal-to-noise
snr_max: 20 # ratio between 10 and 20 dBs
collection: MUSAN.Collection.BackgroundNoise # use background noise from MUSAN
# (needs pyannote.db.musan)
feature_extraction:
name: LibrosaMFCC # use MFCC from librosa
params:
e: False # do not use energy
De: True # use energy 1st derivative
DDe: True # use energy 2nd derivative
coefs: 19 # use 19 MFCC coefficients
D: True # use coefficients 1st derivative
DD: True # use coefficients 2nd derivative
duration: 0.025 # extract MFCC from 25ms windows
step: 0.010 # extract MFCC every 10ms
sample_rate: 16000 # convert to 16KHz first (if needed)
architecture:
name: StackedRNN
params:
instance_normalize: True # normalize sequences
rnn: LSTM # use LSTM (could be GRU)
recurrent: [128, 128] # two layers with 128 hidden states
bidirectional: True # bidirectional LSTMs
linear: [32, 32] # add two linear layers at the end
scheduler:
name: CyclicScheduler # use cyclic learning rate (LR) scheduler
params:
learning_rate: auto # automatically guess LR upper bound
epochs_per_cycle: 14 # 14 epochs per cycle
The following command will train the network using the training set of AMI database for 1000 epochs:
$ export EXPERIMENT_DIR=tutorials/models/speaker_change_detection
$ pyannote-change-detection train --gpu --to=1000 ${EXPERIMENT_DIR} AMI.SpeakerDiarization.MixHeadset
This will create a bunch of files in TRAIN_DIR
(defined below).
One can follow along the training process using tensorboard.
$ tensorboard --logdir=${EXPERIMENT_DIR}
To get a quick idea of how the network is doing during training, one can use the validate
mode.
It can (should!) be run in parallel to training and evaluates the model epoch after epoch.
One can use tensorboard to follow the validation process.
$ export TRAIN_DIR=${EXPERIMENT_DIR}/train/AMI.SpeakerDiarization.MixHeadset.train
$ pyannote-change-detection validate --purity=0.8 ${TRAIN_DIR} AMI.SpeakerDiarization.MixHeadset
In practice, it is tuning a simple speaker change detection pipeline (pyannote.audio.pipeline.speaker_change_detection.SpeakerChangeDetection) after each epoch and stores the best hyper-parameter configuration on disk:
$ cat ${TRAIN_DIR}/validate/AMI.SpeakerDiarization.MixHeadset/params.yml
epoch: 870
params:
alpha: 0.17578125
min_duration: 0.0
One can also use tensorboard to follow the validation process.
Now that we know how the model is doing, we can apply it on all files of the AMI database and store raw change scores in /path/to/precomputed/scd
:
$ pyannote-change-detection apply ${TRAIN_DIR}/weights/0870.pt AMI.SpeakerDiarization.MixHeadset /path/to/precomputed/scd
We can then use these raw scores to perform actual speaker change detection, and pyannote.metrics
to evaluate the result:
# AMI protocol
>>> from pyannote.database import get_protocol
>>> protocol = get_protocol('AMI.SpeakerDiarization.MixHeadset')
# precomputed scores
>>> from pyannote.audio.features import Precomputed
>>> precomputed = Precomputed('/path/to/precomputed/scd')
# peak detection
>>> from pyannote.audio.signal import Peak
# alpha / min_duration are tunable parameters (and should be tuned for better performance)
# we use log_scale = True because of the final log-softmax in the StackedRNN model
>>> peak = Peak(alpha=0.17, min_duration=0.0, log_scale=True)
# evaluation metric
>>> from pyannote.metrics.diarization import DiarizationPurityCoverageFMeasure
>>> metric = DiarizationPurityCoverageFMeasure()
# loop on test files
>>> from pyannote.database import get_annotated
>>> for test_file in protocol.test():
... # load reference annotation
... reference = test_file['annotation']
... uem = get_annotated(test_file)
...
... # load precomputed change scores as pyannote.core.SlidingWindowFeature
... scd_scores = precomputed(test_file)
...
... # binarize scores to obtain speech regions as pyannote.core.Timeline
... hypothesis = peak.apply(scd_scores, dimension=1)
...
... # evaluate speech activity detection
... metric(reference, hypothesis.to_annotation(), uem=uem)
>>> purity, coverage, fmeasure = metric.compute_metrics()
>>> print(f'Purity = {100*purity:.1f}% / Coverage = {100*coverage:.1f}%')
For more options, see:
$ pyannote-change-detection --help
That's all folks!