Ensure that clusterless speakers go last in the binary diarization

pyannote · Nov 14, 2023 · d1be399 · d1be399
1 parent 05a333c
commit d1be399
Showing 1 changed file with 3 additions and 1 deletion.
diff --git a/pyannote/audio/pipelines/utils/diarization.py b/pyannote/audio/pipelines/utils/diarization.py
@@ -230,7 +230,9 @@ def to_diarization(
         activations = activations.crop(extent, return_data=False)
         count = count.crop(extent, return_data=False)
 
-        sorted_speakers = np.argsort(-activations, axis=-1)
+        # stable sort ensures that all the padding speakers (ones which were added
+        # above via padding based on `count` maximum) are always in the end of speaker list.
+        sorted_speakers = np.argsort(-activations, axis=-1, kind="stable")
         binary = np.zeros_like(activations.data)
 
         for t, ((_, c), speakers) in enumerate(zip(count, sorted_speakers)):