def preprocess(self, current_file, identifier=None):
"""Pre-compute file-wise X and y"""
# extract features for the whole file
# (if it has not been done already)
current_file = self.periodic_preprocess(current_file,
identifier=identifier)
# if labels have already been extracted, do nothing
if identifier in self.preprocessed_.setdefault('y', {}):
return current_file
# get features as pyannote.core.SlidingWindowFeature instance
X = self.preprocessed_['X'][identifier]
sw = X.sliding_window
n_samples = X.getNumber()
y = np.zeros((n_samples + 4, 1), dtype=np.int8) - 1
# [-1] ==> unknown / [0] ==> not change part / [1] ==> change part
annotated = get_annotated(current_file)
annotation = current_file['annotation']
segments = []
for segment, _ in annotation.itertracks():
segments.append(
Segment(segment.start - self.balance,
segment.start + self.balance))
segments.append(
Segment(segment.end - self.balance,
segment.end + self.balance))
change_part = Timeline(segments).support().crop(annotated,
mode='intersection')
# iterate over non-change regions
for non_changes in change_part.gaps(annotated):
indices = sw.crop(non_changes, mode='loose')
y[indices, 0] = 0
# iterate over change regions
for changes in change_part:
indices = sw.crop(changes, mode='loose')
y[indices, 0] = 1
y = SlidingWindowFeature(y[:-1], sw)
self.preprocessed_['y'][identifier] = y
return current_file
def main():
usage = "%prog [options] RTTMone RTTMtwo"
desc = "Convert the txtfile from diarization of the from: \
ID t_in t_out \
into a kaldi format file for spkdet task"
version = "%prog 0.1"
parser = OptionParser(usage=usage, description=desc, version=version)
(opt, args) = parser.parse_args()
if (len(args) != 3):
parser.error("Incorrect number of arguments")
vadrttm, overlaprttm, outputrttm = args
# Read document and loaded in memory
vad = pyannote.database.util.load_rttm(vadrttm)
ovl = pyannote.database.util.load_rttm(overlaprttm)
fw = open(outputrttm, 'wt')
for name in vad:
# Examples
# speech = vad['EN2002a.Mix-Headset-0000000-0006000'].get_timeline()
# duration = vad['EN2002a.Mix-Headset-0000000-0006000'].get_timeline()[-1][1]
# overlap = ovl['EN2002a.Mix-Headset-0000000-0006000'].get_timeline()
speech = vad[name].get_timeline()
duration = vad[name].get_timeline()[-1][1]
if name in ovl.keys():
overlap = ovl[name].get_timeline()
# just get the intersections of the VAD and overlap
intersection = Timeline()
for speech_segment, overlap_segment in speech.co_iter(overlap):
intersection.add(speech_segment & overlap_segment)
keep = intersection.gaps(support=Segment(0, duration))
vad_without_overlap = speech.crop(keep)
else:
vad_without_overlap = speech
# Write RTTM
write_rttm(fw, vad_without_overlap, label='speech')
fw.close()
def run(self):
# wav file duration
wav = self.in_wav().path
with contextlib.closing(wave.open(wav, 'r')) as f:
frames = f.getnframes()
rate = f.getframerate()
duration = frames / rate
extent = Segment(0., duration)
with self.in_speaker().open('r') as fp:
speaker = pyannote.core.json.load(fp)
timeline = Timeline()
for segment, _ in speaker.itertracks():
timeline.add(segment)
# fill gaps
for gap in timeline.gaps(extent):
if gap.duration < self.fill_gaps:
timeline.add(gap)
timeline = timeline.coverage()
# dump as annotation...
if self.to_annotation:
annotation = Annotation()
for s, segment in enumerate(timeline):
annotation[segment] = s
annotation = annotation.anonymize_labels(generator='string')
with self.out_put().open('w') as fp:
pyannote.core.json.dump(annotation, fp)
# ... or as timeline
else:
with self.out_put().open('w') as fp:
pyannote.core.json.dump(timeline, fp)
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 serial_speaker_to_Annotation(serial_speaker, uri=None, modality='speaker'):
"""
Parameters:
-----------
serial_speaker : `dict`
loaded from a serial speaker JSON as defined
in https://figshare.com/articles/TV_Series_Corpus/3471839
uri (uniform resource identifier) : `str`, optional
which identifies the annotation (e.g. episode number)
Default : None
modality : `str`, optional
modality of the annotation as defined in https://github.com/pyannote/pyannote-core
Returns:
--------
annotation: pyannote `Annotation`
for speaker identification/diarization as defined
in https://github.com/pyannote/pyannote-core
annotated: pyannote `Timeline`
representing the annotated parts of the serial_speaker file
Unknown speakers are not considered as annotated
"""
annotation = Annotation(uri, modality)
not_annotated = Timeline(uri=uri)
for segment in serial_speaker["data"]["speech_segments"]:
time = Segment(segment["start"], segment["end"])
speaker_id = segment['speaker'].replace(" ", "_")
annotation[time, speaker_id] = speaker_id
if speaker_id == 'unknown':
not_annotated.add(time)
end = serial_speaker.get("duration", segment["end"])
annotated = not_annotated.gaps(support=Segment(0.0, end))
return annotation, annotated
def apply(self, predictions, dimension=0):
"""
Parameters
----------
predictions : SlidingWindowFeature
Must be mono-dimensional
dimension : int, optional
Which dimension to process
"""
if len(predictions.data.shape) == 1:
data = predictions.data
elif predictions.data.shape[1] == 1:
data = predictions.data[:, 0]
else:
data = predictions.data[:, dimension]
if self.log_scale:
data = np.exp(data)
n_samples = predictions.getNumber()
window = predictions.sliding_window
timestamps = [window[i].middle for i in range(n_samples)]
# initial state
start = timestamps[0]
label = data[0] > self.onset
if self.scale == 'absolute':
mini = 0
maxi = 1
elif self.scale == 'relative':
mini = np.nanmin(data)
maxi = np.nanmax(data)
elif self.scale == 'percentile':
mini = np.nanpercentile(data, 1)
maxi = np.nanpercentile(data, 99)
onset = mini + self.onset * (maxi - mini)
offset = mini + self.offset * (maxi - mini)
# timeline meant to store 'active' segments
active = Timeline()
for t, y in zip(timestamps[1:], data[1:]):
# currently active
if label:
# switching from active to inactive
if y < offset:
segment = Segment(start - self.pad_onset,
t + self.pad_offset)
active.add(segment)
start = t
label = False
# currently inactive
else:
# switching from inactive to active
if y > onset:
start = t
label = True
# if active at the end, add final segment
if label:
segment = Segment(start - self.pad_onset, t + self.pad_offset)
active.add(segment)
# because of padding, some 'active' segments might be overlapping
# therefore, we merge those overlapping segments
active = active.support()
# remove short 'active' segments
active = Timeline(
[s for s in active if s.duration > self.min_duration_on])
# fill short 'inactive' segments
inactive = active.gaps()
for s in inactive:
if s.duration < self.min_duration_off:
active.add(s)
active = active.support()
return active
def gecko_JSON_to_Annotation(gecko_JSON,
uri=None,
modality='speaker',
confidence_threshold=0.0,
collar=0.0,
expected_min_speech_time=0.0,
manual=False):
"""
Parameters:
-----------
gecko_JSON : `dict`
loaded from a Gecko-compliant JSON as defined in xml_to_GeckoJSON
uri (uniform resource identifier) : `str`
which identifies the annotation (e.g. episode number)
Default : None
modality : `str`
modality of the annotation as defined in https://github.com/pyannote/pyannote-core
confidence_threshold : `float`, Optional.
The segments with confidence under confidence_threshold won't be added to UEM file.
Defaults to keep every segment (i.e. 0.0)
collar: `float`, Optional.
Merge tracks with same label and separated by less than `collar` seconds.
Defaults to keep tracks timeline untouched (i.e. 0.0)
expected_min_speech_time: `float`, Optional.
Threshold (in seconds) under which the total duration of speech time is suspicious (warns the user).
Defaults to never suspect anything (i.e. 0.0)
manual : `bool`
Whether the json is coming from a manual correction or straight from
the forced-alignment output.
In the former case, the regions timing is used. `confidence_threshold`
and `collar` are thus irrelevant.
In the latter case (default), the timing of each term is used.
Returns:
--------
annotation: pyannote `Annotation`
for speaker identification/diarization as defined in https://github.com/pyannote/pyannote-core
annotated: pyannote `Timeline`
representing the annotated parts of the gecko_JSON files (depends on confidence_threshold)
"""
annotation = Annotation(uri, modality)
not_annotated = Timeline(uri=uri)
for monologue in gecko_JSON["monologues"]:
if not monologue:
continue
# '@' defined in https://github.com/hbredin/pyannote-db-plumcot/blob/develop/CONTRIBUTING.md#idepisodetxt
# '+' defined in https://github.com/gong-io/gecko/blob/master/app/geckoModule/constants.js#L35
speaker_ids = re.split("@|\+", monologue["speaker"]["id"])
if manual:
for speaker_id in speaker_ids: # most of the time there's only one
if speaker_id != '': # happens with "all@"
annotation[Segment(monologue["start"], monologue["end"]),
speaker_id] = speaker_id
else:
for i, term in enumerate(monologue["terms"]):
for speaker_id in speaker_ids: # most of the time there's only one
if speaker_id != '': # happens with "all@"
annotation[Segment(term["start"], term["end"]),
speaker_id] = speaker_id
if term["confidence"] <= confidence_threshold:
not_annotated.add(Segment(term["start"], term["end"]))
if manual:
annotated = Timeline([Segment(0.0, monologue["end"])], uri)
else:
annotation = annotation.support(collar)
annotated = not_annotated.gaps(support=Segment(0.0, term["end"]))
total_speech_time = annotation.crop(annotated).get_timeline().duration()
if total_speech_time < expected_min_speech_time:
warnings.warn(
f"total speech time of {uri} is only {total_speech_time})")
return annotation, annotated
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 apply(self, predictions, dimension=0):
"""
Parameters
----------
predictions : SlidingWindowFeature
Must be mono-dimensional
dimension : int, optional
Which dimension to process
"""
if len(predictions.data.shape) == 1:
data = predictions.data
elif predictions.data.shape[1] == 1:
data = predictions.data[:, 0]
else:
data = predictions.data[:, dimension]
n_samples = predictions.getNumber()
window = predictions.sliding_window
timestamps = [window[i].middle for i in range(n_samples)]
# initial state
start = timestamps[0]
label = data[0] > self.onset
# timeline meant to store 'active' segments
active = Timeline()
for t, y in zip(timestamps[1:], data[1:]):
# currently active
if label:
# switching from active to inactive
if y < self.offset:
segment = Segment(start - self.pad_onset,
t + self.pad_offset)
active.add(segment)
start = t
label = False
# currently inactive
else:
# switching from inactive to active
if y > self.onset:
start = t
label = True
# if active at the end, add final segment
if label:
segment = Segment(start - self.pad_onset, t + self.pad_offset)
active.add(segment)
# because of padding, some 'active' segments might be overlapping
# therefore, we merge those overlapping segments
active = active.coverage()
# remove short 'active' segments
active = Timeline(
[s for s in active if s.duration > self.min_duration[1]])
# fill short 'inactive' segments
inactive = active.gaps()
for s in inactive:
if s.duration < self.min_duration[0]:
active.add(s)
active = active.coverage()
return active
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()
def test_empty_gaps():
empty_timeline = Timeline(uri='MyEmptyGaps')
assert list(empty_timeline.gaps()) == []
Segment.set_precision(3)
assert list(empty_timeline.gaps()) == []
def apply(self, predictions, dimension=0):
"""
Parameters
----------
predictions : SlidingWindowFeature
Must be mono-dimensional
dimension : int, optional
Which dimension to process
"""
if len(predictions.data.shape) == 1:
data = predictions.data
elif predictions.data.shape[1] == 1:
data = predictions.data[:, 0]
else:
data = predictions.data[:, dimension]
if self.log_scale:
data = np.exp(data)
n_samples = predictions.getNumber()
window = predictions.sliding_window
timestamps = [window[i].middle for i in range(n_samples)]
# initial state
start = timestamps[0]
label = data[0] > self.onset
if self.scale == 'absolute':
mini = 0
maxi = 1
elif self.scale == 'relative':
mini = np.nanmin(data)
maxi = np.nanmax(data)
elif self.scale == 'percentile':
mini = np.nanpercentile(data, 1)
maxi = np.nanpercentile(data, 99)
onset = mini + self.onset * (maxi - mini)
offset = mini + self.offset * (maxi - mini)
# timeline meant to store 'active' segments
active = Timeline()
for t, y in zip(timestamps[1:], data[1:]):
# currently active
if label:
# switching from active to inactive
if y < offset:
segment = Segment(start - self.pad_onset,
t + self.pad_offset)
active.add(segment)
start = t
label = False
# currently inactive
else:
# switching from inactive to active
if y > onset:
start = t
label = True
# if active at the end, add final segment
if label:
segment = Segment(start - self.pad_onset, t + self.pad_offset)
active.add(segment)
# because of padding, some 'active' segments might be overlapping
# therefore, we merge those overlapping segments
active = active.support()
# remove short 'active' segments
active = Timeline(
[s for s in active if s.duration > self.min_duration_on])
# fill short 'inactive' segments
inactive = active.gaps()
for s in inactive:
if s.duration < self.min_duration_off:
active.add(s)
active = active.support()
return active
