mirror of
https://github.com/pyannote/pyannote-audio.git
synced 2026-11-20 13:07:51 +09:00
Merge branch 'develop' into v2
This commit is contained in:
6 files changed
+52
-12
No files matched your search
@@ -1,4 +1,3 @@
|
||||
|
||||
##############
|
||||
pyannote.audio
|
||||
##############
|
||||
|
||||
@@ -3,7 +3,7 @@
|
||||
|
||||
# The MIT License (MIT)
|
||||
|
||||
# Copyright (c) 2017-2019 CNRS
|
||||
# Copyright (c) 2017-2020 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
|
||||
|
||||
@@ -34,4 +34,8 @@ from .speech_activity_detection import SpeechActivityDetection
|
||||
from .overlap_detection import OverlapDetection
|
||||
from .speech_turn_segmentation import SpeechTurnSegmentation
|
||||
from .speaker_diarization import SpeakerDiarization
|
||||
from .resegmentation import Resegmentation
|
||||
try:
|
||||
from .resegmentation import Resegmentation
|
||||
except ModuleNotFoundError as e:
|
||||
msg = f'Resegmentation module is not available: {e}'
|
||||
print(msg)
|
||||
@@ -32,6 +32,7 @@ from pyannote.core import Annotation
|
||||
from pyannote.database import get_annotated
|
||||
|
||||
from pyannote.metrics.diarization import GreedyDiarizationErrorRate
|
||||
from pyannote.metrics.diarization import DiarizationPurityCoverageFMeasure
|
||||
|
||||
from .speech_turn_segmentation import SpeechTurnSegmentation
|
||||
from .speech_turn_clustering import SpeechTurnClustering
|
||||
@@ -42,7 +43,6 @@ from typing import Union
|
||||
from pyannote.pipeline import Pipeline
|
||||
from pyannote.pipeline.parameter import Uniform
|
||||
|
||||
|
||||
class SpeakerDiarization(Pipeline):
|
||||
"""Speaker diarization pipeline
|
||||
|
||||
@@ -61,7 +61,10 @@ class SpeakerDiarization(Pipeline):
|
||||
Clustering method. Defaults to 'pool'.
|
||||
evaluation_only : `bool`
|
||||
Only process the evaluated regions. Default to False.
|
||||
|
||||
purity : `float`, optional
|
||||
Optimize coverage for target purity.
|
||||
Defaults to optimizing diarization error rate.
|
||||
|
||||
Hyper-parameters
|
||||
----------------
|
||||
min_duration : `float`
|
||||
@@ -74,16 +77,17 @@ class SpeakerDiarization(Pipeline):
|
||||
embedding: Optional[Path] = None,
|
||||
metric: Optional[str] = 'cosine',
|
||||
method: Optional[str] = 'pool',
|
||||
evaluation_only: Optional[bool] = False):
|
||||
evaluation_only: Optional[bool] = False,
|
||||
purity: Optional[float] = None):
|
||||
|
||||
super().__init__()
|
||||
|
||||
self.sad_scores = sad_scores
|
||||
self.scd_scores = scd_scores
|
||||
self.speech_turn_segmentation = SpeechTurnSegmentation(
|
||||
sad_scores=self.sad_scores,
|
||||
scd_scores=self.scd_scores)
|
||||
self.evaluation_only = evaluation_only
|
||||
self.purity = purity
|
||||
|
||||
self.min_duration = Uniform(0, 10)
|
||||
|
||||
@@ -161,10 +165,43 @@ class SpeakerDiarization(Pipeline):
|
||||
|
||||
# TODO. add GMM-based resegmentation
|
||||
|
||||
def loss(self, current_file: dict, hypothesis: Annotation) -> float:
|
||||
"""Compute (1 - coverage) at target purity
|
||||
|
||||
If purity < target, return 1 + (1 - purity)
|
||||
|
||||
Parameters
|
||||
----------
|
||||
current_file : `dict`
|
||||
File as provided by a pyannote.database protocol.
|
||||
hypothesis : `pyannote.core.Annotation`
|
||||
Speech turns.
|
||||
|
||||
Returns
|
||||
-------
|
||||
loss : `float`
|
||||
1. - cluster coverage.
|
||||
"""
|
||||
|
||||
metric = DiarizationPurityCoverageFMeasure()
|
||||
reference = current_file['annotation']
|
||||
uem = get_annotated(current_file)
|
||||
f_measure = metric(reference, hypothesis, uem=uem)
|
||||
purity, coverage, _ = metric.compute_metrics()
|
||||
if purity > self.purity:
|
||||
return 1. - coverage
|
||||
else:
|
||||
return 1. + (1. - purity)
|
||||
|
||||
def get_metric(self) -> GreedyDiarizationErrorRate:
|
||||
"""Return new instance of diarization error rate metric"""
|
||||
return GreedyDiarizationErrorRate(collar=0.0, skip_overlap=False)
|
||||
|
||||
|
||||
# defaults to optimizing diarization error rate
|
||||
if self.purity is None:
|
||||
return GreedyDiarizationErrorRate(collar=0.0, skip_overlap=False)
|
||||
|
||||
# fallbacks to using self.loss(...)
|
||||
raise NotImplementedError()
|
||||
|
||||
class Yin2018(SpeakerDiarization):
|
||||
"""Speaker diarization pipeline introduced in Yin et al., 2018
|
||||
|
||||
+2
-2
@@ -1,8 +1,8 @@
|
||||
cachetools >= 2.0.0
|
||||
librosa >= 0.7.0
|
||||
pandas >= 0.18.0
|
||||
pyannote.core >= 3.2
|
||||
pyannote.database >= 2.4.1
|
||||
pyannote.core >= 3.3
|
||||
pyannote.database >= 2.4.2
|
||||
pyannote.generators >= 2.0
|
||||
pyannote.metrics >= 2.2
|
||||
pyannote.pipeline >= 1.2
|
||||
|
||||
@@ -172,7 +172,7 @@ It can be run while the model is still training and evaluates the model every 20
|
||||
In practice, it is tuning a simple speaker verification experiment and stores the best hyper-parameter configuration on disk:
|
||||
|
||||
```bash
|
||||
$ export VAL_DIR = ${TRN_DIR}/validate/VoxCeleb.SpeakerDiarization.VoxCeleb1.test
|
||||
$ export VAL_DIR = ${TRN_DIR}/validate_equal_error_rate/VoxCeleb.SpeakerDiarization.VoxCeleb1.test
|
||||
$ cat ${VAL_DIR}/params.yml
|
||||
```
|
||||
```yaml
|
||||
|
||||
Reference in new issue
Block a user