The MIT License (MIT)
Copyright (c) 2019-2020 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
This tutorial assumes that you have already followed the data preparation tutorial.
For the purpose of this tutorial, we use models available on torch.hub that were pretrained on AMI training subset:
import torch
# speech activity detection model trained on AMI training set
sad = torch.hub.load('pyannote/pyannote-audio', 'sad_ami')
# speaker change detection model trained on AMI training set
scd = torch.hub.load('pyannote/pyannote-audio', 'scd_ami')
# overlapped speech detection model trained on AMI training set
ovl = torch.hub.load('pyannote/pyannote-audio', 'ovl_ami')
# speaker embedding model trained on AMI training set
emb = torch.hub.load('pyannote/pyannote-audio', 'emb_ami')Note that models will run on GPU by default if one is available.
Both device and batch size can be specified manually if needed:
sad = torch.hub.load('pyannote/pyannote-audio', 'sad_ami', device='cpu', batch_size=128)We will apply those pretrained models on the first file of the AMI test subset.
# ... or use a file provided by a pyannote.database protocol
# in this example, we are using AMI first test file.
from pyannote.database import get_protocol
from pyannote.database import FileFinder
preprocessors = {'audio': FileFinder()}
protocol = get_protocol('AMI.SpeakerDiarization.MixHeadset',
preprocessors=preprocessors)
test_file = next(protocol.test())Pretrained class:
from pyannote.audio.features import Pretrained
sad = Pretrained(validate_dir='/path/to/validation/directory')# one can use their own file like this...
test_file = {'uri': 'filename', 'audio': '/path/to/your/filename.wav'}Note that, in case of domain mismatch between your data and the AMI corpus, you might be better off training your own models or fine-tuning a pretrained one.
# obtain raw SAD scores (as `pyannote.core.SlidingWindowFeature` instance)
sad_scores = sad(test_file)
# binarize raw SAD scores
# NOTE: both onset/offset values were tuned on AMI dataset.
# you might need to use different values for better results.
from pyannote.audio.utils.signal import Binarize
binarize = Binarize(offset=0.52, onset=0.52, log_scale=True,
min_duration_off=0.1, min_duration_on=0.1)
# speech regions (as `pyannote.core.Timeline` instance)
speech = binarize.apply(sad_scores, dimension=1)# obtain raw SCD scores (as `pyannote.core.SlidingWindowFeature` instance)
scd_scores = scd(test_file)
# detect peaks and return speaker homogeneous segments
# NOTE: both alpha/min_duration values were tuned on AMI dataset.
# you might need to use different values for better results.
from pyannote.audio.utils.signal import Peak
peak = Peak(alpha=0.10, min_duration=0.10, log_scale=True)
# speaker change point (as `pyannote.core.Timeline` instance)
partition = peak.apply(scd_scores, dimension=1)# obtain raw OVL scores (as `pyannote.core.SlidingWindowFeature` instance)
ovl_scores = ovl(test_file)
# binarize raw OVL scores
# NOTE: both onset/offset values were tuned on AMI dataset.
# you might need to use different values for better results.
from pyannote.audio.utils.signal import Binarize
binarize = Binarize(offset=0.55, onset=0.55, log_scale=True,
min_duration_off=0.1, min_duration_on=0.1)
# overlapped speech regions (as `pyannote.core.Timeline` instance)
overlap = binarize.apply(ovl_scores, dimension=1)# let's visualize SAD, SCD and OVL results using pyannote.core visualization API
import numpy as np
from matplotlib import pyplot as plt
from pyannote.core import Segment, notebook
# only plot one minute (between t=120s and t=180s)
notebook.crop = Segment(120, 180)
# helper function to make visualization prettier
from pyannote.core import SlidingWindowFeature
plot_ready = lambda scores: SlidingWindowFeature(np.exp(scores.data[:, 1:]), scores.sliding_window)
# create a figure with 8 rows with matplotlib
nrows = 8
fig, ax = plt.subplots(nrows=nrows, ncols=1)
fig.set_figwidth(20)
fig.set_figheight(nrows * 2)
# 1st row: reference annotation
notebook.plot_annotation(test_file['annotation'], ax=ax[0], time=False)
ax[0].text(notebook.crop.start + 0.5, 0.1, 'reference', fontsize=14)
# 2nd row: SAD raw scores
notebook.plot_feature(plot_ready(sad_scores), ax=ax[1], time=False)
ax[1].text(notebook.crop.start + 0.5, 0.1, 'speech activity\ndetection scores', fontsize=14)
ax[1].set_ylim(-0.1, 1.1)
# 3rd row: SAD result
notebook.plot_timeline(speech, ax=ax[2], time=False)
ax[2].text(notebook.crop.start + 0.5, 0.1, 'speech activity detection', fontsize=14)
# 4th row: SCD raw scores
notebook.plot_feature(plot_ready(scd_scores), ax=ax[3], time=False)
ax[3].text(notebook.crop.start + 0.5, 0.1, 'speaker change\ndetection scores', fontsize=14)
ax[3].set_ylim(-0.1, 0.6)
# 5th row: SCD result
notebook.plot_timeline(partition, ax=ax[4], time=False)
ax[4].text(notebook.crop.start + 0.5, 0.1, 'speaker change detection', fontsize=14)
# 6th row: OVL raw scores
notebook.plot_feature(plot_ready(ovl_scores), ax=ax[5], time=False)
ax[5].text(notebook.crop.start + 0.5, 0.2, 'overlapped speech\ndetection scores', fontsize=14)
ax[5].set_ylim(-0.1, 1.1)
# 7th row: OVL result
notebook.plot_timeline(overlap, ax=ax[6], time=False)
ax[6].text(notebook.crop.start + 0.5, 0.1, 'overlapped speech detection', fontsize=14)
# 8th row: reference annotation
notebook.plot_annotation(test_file['annotation'], ax=ax[7], legend=False)
_ = ax[7].text(notebook.crop.start + 0.5, 0.1, 'reference', fontsize=14)
# speech turns are simply the intersection of SAD and SCD
speech_turns = partition.crop(speech)# obtain raw embeddings (as `pyannote.core.SlidingWindowFeature` instance)
embeddings = emb(test_file)
chunks = embeddings.sliding_window
print(f'Embeddings were extracted every {1000 * chunks.step:g}ms on {1000 * chunks.duration:g}ms-long windows.')# for the purpose of this tutorial, we only work of long (> 1s) speech turns
from pyannote.core import Timeline
long_turns = Timeline(segments=[s for s in speech_turns if s.duration > 2.])Extracting embedding for a given speech turn is as easy as embedding.crop(segment):
X, Y = [], []
for segment in long_turns:
# "strict" only keeps embedding strictly included in segment
x = embeddings.crop(segment, mode='strict')
# average speech turn embedding
X.append(np.mean(x, axis=0))
# keep track of speaker label (for later scatter plot)
y = test_file['annotation'].argmax(segment)
Y.append(y)
X = np.vstack(X)
_, y_true = np.unique(Y, return_inverse=True)We can use tSNE to visualize (and later cluster, maybe?) embeddings.
# apply tSNE on embeddings
from sklearn.manifold import TSNE
tsne = TSNE(n_components=2, metric="cosine")
X_2d = tsne.fit_transform(X)
# plot
fig, ax = plt.subplots()
fig.set_figheight(5)
fig.set_figwidth(5)
plt.scatter(*X_2d.T, c=y_true)
That's all folks!