def test_remove_and_extent():
t = Timeline(uri='MyAudioFile')
t.add(Segment(6, 8))
t.add(Segment(7, 9))
t.add(Segment(6, 9))
t.remove(Segment(6, 9))
assert t.extent() == Segment(6, 9)
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 test_update_extent():
timeline = Timeline([Segment(0, 1), Segment(2, 3), Segment(4, 5)])
other_timeline = Timeline([Segment(1.5, 6)])
timeline.update(other_timeline)
assert timeline.extent() == Segment(0, 6)
def trn_iter(self):
# the train set comes from Fisher, Switchboard and Mixer NIST SRE datasets
# the list for training has no speaker's turn annotations, since every utterance has one speaker only
# the provided file list also requires special care for Idiap location, since the folder structure is different
extend_uri = None
if os.path.isdir("/idiap") and "USER" in os.environ:
extend_uri = {
'fishe1': 8,
'fishe2': 8,
'mix05': 2,
'mix06': 2,
'swcellp1': 5,
'swcellp2': 5,
'swph2': 5,
'swph3': 5
}
# load UEM list, which was generated by recording the duration of each utterance
data_dir = Path(__file__).parent / 'data' / 'speaker_diarization'
annotated = data_dir / f'train-world.uem'
names = ['uri', 'NA0', 'start', 'end']
annotated = read_table(annotated, delim_whitespace=True, names=names)
# load the list of utterances with speaker_ids, it's not an MDTM file
annotation = data_dir / f'train-world.lst'
names = ['uri', 'speaker']
annotation = read_table(annotation, delim_whitespace=True, names=names)
AnnotatedGroups = annotated.groupby(by='uri')
AnnotationGroups = annotation.groupby(by='uri')
samples_with_annotations = AnnotationGroups.groups.keys()
for raw_uri, annotated in AnnotatedGroups:
# process only those uris that we have inside annotations
if raw_uri in samples_with_annotations:
# from raw_uri, construct correct path to the audio file
uri = raw_uri
if extend_uri is not None:
uriparts = raw_uri.split('/')
if uriparts[0] in extend_uri.keys():
uri = os.path.join(
uriparts[0], uriparts[1][:extend_uri[uriparts[0]]],
uriparts[1])
segments = []
for segment in annotated.itertuples():
segments.append(
Segment(start=segment.start, end=segment.end))
timeline = Timeline(uri=uri, segments=segments)
annotation = Annotation(uri=uri)
for t, turn in enumerate(
AnnotationGroups.get_group(raw_uri).itertuples()):
segment = timeline.extent()
annotation[segment, t] = turn.speaker
current_file = {
'database': 'SRE',
'uri': uri,
'annotated': timeline,
'annotation': annotation
}
yield current_file
