This repository is no longer maintained: a complete rewrite (leading to much better performance) is now part of pyannote-audio, which provides:

a tutorial to train, validate, and apply a TristouNet embedding
a pre-trained TristouNet model

TristouNet: Triplet Loss for Speaker Turn Embedding

Code for http://arxiv.org/abs/1609.04301

Citation

@inproceedings{Bredin2017,
    author = {Herv\'{e} Bredin},
    title = {{TristouNet: Triplet Loss for Speaker Turn Embedding}},
    booktitle = {42nd IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2017},
    year = {2017},
    url = {http://arxiv.org/abs/1609.04301},
}

Installation

Foreword: Sorry about the python=2.7 constraint but yaafe does not support Python 3 at the time of writing this.
Ping me if this is no longer true!

$ conda create --name tristounet python=2.7 anaconda
$ source activate tristounet
$ conda install gcc
$ conda install -c yaafe yaafe=0.65
$ pip install "theano==0.8.2"
$ pip install "keras==1.1.0"
$ pip install "pyannote.metrics==0.9"
$ pip install "pyannote.audio==0.1.4"
$ pip install "pyannote.db.etape==0.2.1"
$ pip install "pyannote.database==0.4.1"
$ pip install "pyannote.generators==0.1.1"

Using pip freeze, make sure you have installed these exact versions before submitting an issue.

What did I just install?

keras (and its theano backend) is used for all things deep.
yaafe is used for MFCC feature extraction in pyannote.audio. You might also want to checkout librosa (easy to install, but much slower) though pyannote.audio does not support it yet.
pyannote.audio is where the magic happens (TristouNet architecture, triplet loss, and triplet sampling)
pyannote.db.etape is the ETAPE plugin for pyannote.database, a common API for multimedia databases and experimental protocols (e.g. train/dev/test sets definition).
pyannote.metrics provides evaluation metrics.

Then, edit ~/.keras/keras.json to configure keras with theano backend.

$ cat ~/.keras/keras.json
{
    "image_dim_ordering": "th",
    "epsilon": 1e-07,
    "floatx": "float32",
    "backend": "theano"
}

About the ETAPE database

To reproduce the experiment, you obviously need to have access to the ETAPE corpus.
It can be obtained from ELRA catalogue.

However, if you own another corpus with "who speaks when" annotations, then you are just one step away from testing TristouNet on your own corpus.

Just fork pyannote.db.etape and adapt the code to your own database. If (and only if) the database is available for anyone to use, I am willing to help you do that.

Training

For convenience, this script is also available in train.py. See file header for usage instructions.

# environment
>>> WAV_TEMPLATE = '/path/to/where/files/are/stored/{uri}.wav'
>>> LOG_DIR = '/path/to/where/trained/models/are/stored'

# for reproducibility
>>> import numpy as np
>>> np.random.seed(1337)

# feature extraction
>>> from pyannote.audio.features.yaafe import YaafeMFCC
>>> feature_extractor = YaafeMFCC(e=False, De=True, DDe=True,
...                               coefs=11, D=True, DD=True)

# ETAPE database
>>> medium_template = {'wav': WAV_TEMPLATE}
>>> from pyannote.database import Etape
>>> database = Etape(medium_template=medium_template)

# experimental protocol (ETAPE TV subset)
>>> protocol = database.get_protocol('SpeakerDiarization', 'TV')

# TristouNet architecture
>>> from pyannote.audio.embedding.models import TristouNet
>>> architecture = TristouNet()

# triplet loss
>>> from pyannote.audio.embedding.losses import TripletLoss
>>> margin = 0.2    # `alpha` in the paper
>>> loss = TripletLoss(architecture, margin=margin)

>>> from pyannote.audio.embedding.base import SequenceEmbedding
>>> log_dir = LOG_DIR
>>> embedding = SequenceEmbedding(
...     loss=loss, optimizer='rmsprop', log_dir=log_dir)

# sequence duration (in seconds)
>>> duration = 2.0

# triplet sampling
# this might take some time as the whole corpus is loaded in memory,
# and the whole set of MFCC features sequences is precomputed
>>> from pyannote.audio.embedding.generator import TripletBatchGenerator
>>> per_label = 40  # `n` in the paper
>>> batch_size = 32
>>> generator = TripletBatchGenerator(
...     feature_extractor, protocol.train(), embedding, margin=margin,
...     duration=duration, per_label=per_label, batch_size=batch_size)

# shape of feature sequences (n_frames, n_features)
>>> input_shape = generator.get_shape()

# number of samples per epoch
# (rounded to closest batch_size multiple)
>>> samples_per_epoch = per_label * (per_label - 1) * generator.n_labels
>>> samples_per_epoch = samples_per_epoch - (samples_per_epoch % batch_size)

# number of epochs
>>> nb_epoch = 50

# actual training
>>> embedding.fit(input_shape, generator, samples_per_epoch, nb_epoch)

Here is an excerpt of the expected output:

UserWarning: 68 labels (out of 179) have less than 40 training samples.

Epoch 1/1000
278528/278528 [==============================] - 1722s - loss: 0.1074
Epoch 2/1000
278528/278528 [==============================] - 1701s - loss: 0.1179
Epoch 3/1000
278528/278528 [==============================] - 1691s - loss: 0.1204
...
Epoch 48/1000
278528/278528 [==============================] - 1688s - loss: 0.0925
Epoch 49/1000
278528/278528 [==============================] - 1675s - loss: 0.0937
Epoch 50/1000
278528/278528 [==============================] - 1678s - loss: 0.0863

"same/different" toy experiment

For convenience, this script is also available in same_different_experiment.py. See file header for usage instructions.

# for reproducibility
>>> import numpy as np
>>> np.random.seed(1337)

# generate set of labeled sequences
>>> from pyannote.audio.generators.labels import \
...     LabeledFixedDurationSequencesBatchGenerator
>>> generator = LabeledFixedDurationSequencesBatchGenerator(
...     feature_extractor, duration=duration, step=duration, batch_size=-1)
>>> X, y = zip(*generator(protocol.development()))
>>> X, y = np.vstack(X), np.hstack(y)

# randomly select (at most) 100 sequences from each speaker to ensure
# all speakers have the same importance in the evaluation
>>> unique, y, counts = np.unique(y, return_inverse=True, return_counts=True)
>>> n_speakers = len(unique)
>>> indices = []
>>> for speaker in range(n_speakers):
...     i = np.random.choice(np.where(y == speaker)[0], size=min(100, counts[speaker]), replace=False)
...     indices.append(i)
>>> indices = np.hstack(indices)
>>> X, y = X[indices], y[indices, np.newaxis]

# load pre-trained embedding
>>> architecture_yml = log_dir + '/architecture.yml'
>>> weights_h5 = log_dir + '/weights/{epoch:04d}.h5'.format(epoch=nb_epoch - 1)
>>> embedding = SequenceEmbedding.from_disk(architecture_yml, weights_h5)

# embed all sequences
>>> fX = embedding.transform(X, batch_size=batch_size, verbose=0)

# compute euclidean distance between every pair of sequences
>>> from scipy.spatial.distance import pdist
>>> distances = pdist(fX, metric='euclidean')

# compute same/different groundtruth
>>> y_true = pdist(y, metric='chebyshev') < 1

# plot positive/negative scores distribution
# plot DET curve and return equal error rate
>>> from pyannote.metrics.plot.binary_classification import \
...     plot_det_curve, plot_distributions
>>> prefix = log_dir + '/plot.{epoch:04d}'.format(epoch=nb_epoch - 1)
>>> plot_distributions(y_true, distances, prefix, xlim=(0, 2), ymax=3, nbins=100)
>>> eer = plot_det_curve(y_true, -distances, prefix)
>>> print('EER = {eer:.2f}%'.format(eer=100*eer))

This is the expected output:

EER = 14.44%

Baseline results with BIC and Gaussian divergence on the same experimental protocol can be obtained with same_different_experiment_baseline.py. See file header for usage instructions. This is the expected output:

BIC EER = 20.53%
DIV EER = 22.49%

Speaker change detection

For convenience, this script is also available in speaker_change_detection.py. See file header for usage instructions.

# for reproducibility
>>> import numpy as np
>>> np.random.seed(1337)

# load pre-trained embedding
>>> architecture_yml = log_dir + '/architecture.yml'
>>> weights_h5 = log_dir + '/weights/{epoch:04d}.h5'.format(epoch=nb_epoch - 1)
>>> embedding = SequenceEmbedding.from_disk(architecture_yml, weights_h5)

>>> from pyannote.audio.embedding.segmentation import Segmentation
>>> segmentation = Segmentation(embedding, feature_extractor,
...                             duration=duration, step=0.100)

# process files from development set
# (and, while we are at it, load groundtruth for later comparison)
>>> predictions = {}
>>> groundtruth = {}
>>> for test_file in protocol.development():
...     uri = test_file['uri']
...     groundtruth[uri] = test_file['annotation']
...     wav = test_file['medium']['wav']
...     # this is where the magic happens
...     predictions[uri] = segmentation.apply(wav)

# tested thresholds
>>> alphas = np.linspace(0, 1, 50)

# evaluation metrics (purity and coverage)
>>> from pyannote.metrics.segmentation import SegmentationPurity
>>> from pyannote.metrics.segmentation import SegmentationCoverage
>>> purity = [SegmentationPurity() for alpha in alphas]
>>> coverage = [SegmentationCoverage() for alpha in alphas]

# peak detection
>>> from pyannote.audio.signal import Peak
>>> for i, alpha in enumerate(alphas):
...     # initialize peak detection algorithm
...     peak = Peak(alpha=alpha, min_duration=1.0)
...     for uri, reference in groundtruth.items():
...         # apply peak detection
...         hypothesis = peak.apply(predictions[uri])
...         # compute purity and coverage
...         purity[i](reference, hypothesis)
...         coverage[i](reference, hypothesis)

# print the results in three columns:
# threshold, purity, coverage
>>> TEMPLATE = '{alpha:.2f} {purity:.1f}% {coverage:.1f}%'
>>> for i, a in enumerate(alphas):
...     p = 100 * abs(purity[i])
...     c = 100 * abs(coverage[i])
...     print(TEMPLATE.format(alpha=a, purity=p, coverage=c))

Here is an excerpt of the expected result:

0.00 95.2% 38.9%
0.02 95.2% 38.9%
0.04 95.2% 39.0%
...
0.49 94.4% 55.0%
...
0.96 86.7% 93.7%
0.98 86.3% 94.4%
1.00 86.2% 94.9%

Replace Segmentation by BICSegmentation or GaussianDivergenceSegmentation to get baseline results. You should also remove derivatives from MFCC feature extractor (D=False, DD=False, De=False, DDe=False) as it leads to better baseline performance. You might need to install pyannote.algorithms==0.6.5 first.

>>> from pyannote.audio.segmentation import BICSegmentation
>>> segmentation = BICSegmentation(feature_extractor,
...                                covariance_type='full',
...                                duration=duration,
...                                step=0.100)

Here is an excerpt of the expected result:

0.000 95.1% 37.7%
0.020 95.1% 37.7%
0.041 95.1% 37.7%
...
0.388 94.4% 48.0%
...
0.959 86.4% 86.8%
0.980 86.2% 88.0%
1.000 85.8% 88.8%

>>> from pyannote.audio.segmentation import GaussianDivergenceSegmentation
>>> segmentation = GaussianDivergenceSegmentation(feature_extractor,
...                                               duration=duration,
...                                               step=0.100)

Here is an excerpt of the expected result:

0.000 95.2% 36.2%
0.020 95.2% 36.2%
...
0.327 94.4% 48.3%
...
0.959 87.5% 85.8%
0.980 87.4% 86.5%
1.000 87.0% 87.5%

For convenience, two scripts are available: speaker_change_detection_bic.py and speaker_change_detection_div.py. See file header for usage instructions.

hbredin/TristouNet

TristouNet: Triplet Loss for Speaker Turn Embedding

Citation

Installation

About the ETAPE database

Training

"same/different" toy experiment

Speaker change detection