def write_test_file(data_dir, output_file, trial_length):
annotations, max_length, speakers = read_annotaitons(data_dir)
# create an artificial non-overlapping segments each of the trial_length size
trial_segments = Timeline()
for i in range(0, int(max_length) // trial_length):
trial_segments.add(Segment(start=i*trial_length, end=(i+1)*trial_length))
with open(output_file, 'w') as f:
for label in speakers.keys():
for annotation in annotations:
# make sure our trial segments are not extending beyond the total length of the speech data
support = annotation.get_timeline().extent()
# we consider smaller segment here to make sure an embedding of 3 seconds can be computed
adjusted_trial_segments = trial_segments.crop(Segment(start=support.start, end=support.end - 3.),
mode='loose')
uri = annotation.uri
cur_timeline = annotation.label_timeline(label, copy=False)
for trial_segment in adjusted_trial_segments:
cropped_speaker = cur_timeline.crop(trial_segment, mode='intersection')
if not cropped_speaker:
f.write('{0} {1} {2:0>7.2f} {3:0>7.2f} nontarget - -\n'.format(
label,
uri,
trial_segment.start,
trial_segment.end))
else:
f.write('{0} {1} {2:0>7.2f} {3:0>7.2f} target {4:0>7.2f} {5:0>7.2f}\n'.format(
label,
uri,
trial_segment.start,
trial_segment.end,
cropped_speaker[0].start,
cropped_speaker[0].duration))
def test_crop_mapping():
timeline = Timeline([Segment(0, 2), Segment(1, 2), Segment(3, 4)])
cropped, mapping = timeline.crop(Segment(1, 2), returns_mapping=True)
expected_cropped = Timeline([Segment(1, 2)])
assert cropped == expected_cropped
expected_mapping = {Segment(1, 2): [Segment(0, 2), Segment(1, 2)]}
assert mapping == expected_mapping
def test_union():
first_timeline = Timeline([Segment(0, 1), Segment(2, 3), Segment(4, 5)])
second_timeline = Timeline([Segment(1.5, 4.5)])
assert first_timeline.union(second_timeline) == Timeline(
[Segment(0, 1),
Segment(1.5, 4.5),
Segment(2, 3),
Segment(4, 5)])
assert second_timeline.crop(first_timeline) == Timeline(
[Segment(2, 3), Segment(4, 4.5)])
assert list(first_timeline.co_iter(second_timeline)) == [
(Segment(2, 3), Segment(1.5, 4.5)), (Segment(4, 5), Segment(1.5, 4.5))
]
def test_union():
first_timeline = Timeline([Segment(0, 1),
Segment(2, 3),
Segment(4, 5)])
second_timeline = Timeline([Segment(1.5, 4.5)])
assert first_timeline.union(second_timeline) == Timeline([Segment(0, 1),
Segment(1.5, 4.5),
Segment(2, 3),
Segment(4, 5)])
assert second_timeline.crop(first_timeline) == Timeline([Segment(2, 3),
Segment(4, 4.5)])
assert list(first_timeline.co_iter(second_timeline)) == [(Segment(2, 3), Segment(1.5, 4.5)),
(Segment(4, 5), Segment(1.5, 4.5))]
def remove_excluded(self):
if len(self.excluded) == 0:
return
from pyannote.core import Segment, Timeline
segments = []
for recording, _segments in self.segments.groupby(
"recording_filename"):
sampled = Timeline(segments=[
Segment(segment_onset, segment_offset)
for segment_onset, segment_offset in _segments[
["segment_onset", "segment_offset"]].values
])
excl_segments = self.excluded.loc[
self.excluded["recording_filename"] == recording]
excl = Timeline(segments=[
Segment(segment_onset, segment_offset)
for segment_onset, segment_offset in excl_segments[
["segment_onset", "segment_offset"]].values
])
# sampled = sampled.extrude(sampled) # not released yet
extent_tl = Timeline([sampled.extent()], uri=sampled.uri)
truncating_support = excl.gaps(support=extent_tl)
sampled = sampled.crop(truncating_support, mode="intersection")
segments.append(
pd.DataFrame(
[[recording, s.start, s.end] for s in sampled],
columns=[
"recording_filename", "segment_onset", "segment_offset"
],
))
self.segments = pd.concat(segments)
def __call__(self, reference, hypothesis):
if isinstance(reference, Annotation):
reference = reference.get_timeline()
if isinstance(hypothesis, Annotation):
hypothesis = hypothesis.get_timeline()
# over-segmentation
over = Timeline(uri=reference.uri)
prev_r = reference[0]
intersection = []
for r, h in reference.co_iter(hypothesis):
if r != prev_r:
intersection = sorted(intersection)
for _, segment in intersection[:-1]:
over.add(segment)
intersection = []
prev_r = r
segment = r & h
intersection.append((segment.duration, segment))
intersection = sorted(intersection)
for _, segment in intersection[:-1]:
over.add(segment)
# under-segmentation
under = Timeline(uri=reference.uri)
prev_h = hypothesis[0]
intersection = []
for h, r in hypothesis.co_iter(reference):
if h != prev_h:
intersection = sorted(intersection)
for _, segment in intersection[:-1]:
under.add(segment)
intersection = []
prev_h = h
segment = h & r
intersection.append((segment.duration, segment))
intersection = sorted(intersection)
for _, segment in intersection[:-1]:
under.add(segment)
# extent
extent = reference.extent()
# correct (neither under- nor over-segmented)
correct = under.union(over).gaps(support=extent)
# frontier error (both under- and over-segmented)
frontier = under.crop(over)
# under-segmented
not_over = over.gaps(support=extent)
only_under = under.crop(not_over)
# over-segmented
not_under = under.gaps(support=extent)
only_over = over.crop(not_under)
status = Annotation(uri=reference.uri)
# for segment in correct:
# status[segment, '_'] = 'correct'
for segment in frontier:
status[segment, '_'] = 'shift'
for segment in only_over:
status[segment, '_'] = 'over-segmentation'
for segment in only_under:
status[segment, '_'] = 'under-segmentation'
return status.support()
def __call__(self, reference, hypothesis):
if isinstance(reference, Annotation):
reference = reference.get_timeline()
if isinstance(hypothesis, Annotation):
hypothesis = hypothesis.get_timeline()
# over-segmentation
over = Timeline(uri=reference.uri)
prev_r = reference[0]
intersection = []
for r, h in reference.co_iter(hypothesis):
if r != prev_r:
intersection = sorted(intersection)
for _, segment in intersection[:-1]:
over.add(segment)
intersection = []
prev_r = r
segment = r & h
intersection.append((segment.duration, segment))
intersection = sorted(intersection)
for _, segment in intersection[:-1]:
over.add(segment)
# under-segmentation
under = Timeline(uri=reference.uri)
prev_h = hypothesis[0]
intersection = []
for h, r in hypothesis.co_iter(reference):
if h != prev_h:
intersection = sorted(intersection)
for _, segment in intersection[:-1]:
under.add(segment)
intersection = []
prev_h = h
segment = h & r
intersection.append((segment.duration, segment))
intersection = sorted(intersection)
for _, segment in intersection[:-1]:
under.add(segment)
# extent
extent = reference.extent()
# correct (neither under- nor over-segmented)
correct = under.union(over).gaps(focus=extent)
# frontier error (both under- and over-segmented)
frontier = under.crop(over)
# under-segmented
not_over = over.gaps(focus=extent)
only_under = under.crop(not_over)
# over-segmented
not_under = under.gaps(focus=extent)
only_over = over.crop(not_under)
status = Annotation(uri=reference.uri)
for segment in correct:
status[segment, '_'] = 'correct'
for segment in frontier:
status[segment, '_'] = 'frontier'
for segment in only_over:
status[segment, '_'] = 'over'
for segment in only_under:
status[segment, '_'] = 'under'
return status.smooth()