fix: fix alignment between diarization and sources

This commit is contained in:
Lebourdais authored and GitHub committed 2024-10-23 16:50:30 +02:00
1 parent a39991b92d
commit 3f4a22f773
2 files changed
+9 -2

No files matched your search

+1
View File
@@ -11,6 +11,7 @@
### Fixes
- fix: fix clipping issue in speech separation pipeline ([@joonaskalda](https://github.com/joonaskalda/))
- fix: fix alignment between separated sources and diarization when the diarization reference is available ([@Lebourdais](https://github.com/Lebourdais/))
## Version 3.3.2 (2024-09-11)
@@ -124,7 +124,7 @@ class SpeechSeparation(SpeakerDiarizationMixin, Pipeline):
def __init__(
self,
segmentation: PipelineModel = None,
segmentation: PipelineModel = "pyannote/separation-ami-1.0",
segmentation_step: float = 0.1,
embedding: PipelineModel = "speechbrain/spkrec-ecapa-voxceleb@5c0be3875fda05e81f3c004ed8c7c06be308de1e",
embedding_exclude_overlap: bool = False,
@@ -698,6 +698,13 @@ class SpeechSeparation(SpeakerDiarizationMixin, Pipeline):
# strings and integers when reference is available and some hypothesis
# speakers are not present in the reference)
# re-order sources so that they match
# the order given by diarization.labels()
inverse_mapping = {label: index for index, label in mapping.items()}
source.data = sources.data[
:, [inverse_mapping[label] for label in diarization.labels()]
]
if not return_embeddings:
return diarization, sources
@@ -717,7 +724,6 @@ class SpeechSeparation(SpeakerDiarizationMixin, Pipeline):
# re-order centroids so that they match
# the order given by diarization.labels()
inverse_mapping = {label: index for index, label in mapping.items()}
centroids = centroids[
[inverse_mapping[label] for label in diarization.labels()]
]