Update state.py

state.py

@@ -715,19 +715,17 @@ def analyze(inFileName):
         printV("Loaded results", 4)
 
         pipeline = st.session_state.pipeline
-        try:
-            noVoice, oneVoice, multiVoice = su.calcSpeakingTypes(pipeline, currAnnotation, currTotalTime)
-        except Exception as e:
-            print(f"calcSpeakingTypes failed ({e}), falling back to annotation-based voice split")
-            # Fall back: derive voice categories directly from segment timeline.
-            # - oneVoice: each named speaker segment
-            # - noVoice:  gaps between segments (real silence)
-            # - multiVoice: empty (can't detect overlap without the classifier)
+        # For annotation-only files (RTTM/TXT/CSV), annotationToNoiseList
+        # misclassifies almost everything as multiVoice because the window-based
+        # classifier sees >2 speakers in every window. Instead derive voice
+        # categories directly from the annotation's segment gaps — more accurate
+        # and consistent for demo/pre-labeled files.
+        _is_annotation_file = inFileName.lower().endswith((".rttm", ".txt", ".csv"))
+        if _is_annotation_file:
             from pyannote.core import Annotation, Segment
             noVoice    = Annotation()
             multiVoice = Annotation()
             oneVoice   = Annotation()
-            # Sort all segments by start time
             all_segs = sorted(
                 [(seg.start, seg.end, label)
                  for label in currAnnotation.labels()
@@ -735,20 +733,42 @@ def analyze(inFileName):
                  for seg in currAnnotation.subset([label]).itersegments()],
                 key=lambda x: x[0]
             )
-            # Fill oneVoice and detect gaps for noVoice
             prev_end = 0.0
             for start, end, label in all_segs:
                 if start > prev_end + 0.1:
                     noVoice[Segment(prev_end, start)] = 'silence'
                 oneVoice[Segment(start, end)] = label
                 prev_end = max(prev_end, end)
-            # Trailing silence to end of file
             if currTotalTime > prev_end + 0.1:
-                noVoice[Segment(prev_end, currTotalTime)] = 'silence' 
+                noVoice[Segment(prev_end, currTotalTime)] = 'silence'
+        else:
+            try:
+                noVoice, oneVoice, multiVoice = su.calcSpeakingTypes(pipeline, currAnnotation, currTotalTime)
+            except Exception as e:
+                print(f"calcSpeakingTypes failed ({e}), falling back to annotation-based voice split")
+                from pyannote.core import Annotation, Segment
+                noVoice    = Annotation()
+                multiVoice = Annotation()
+                oneVoice   = Annotation()
+                all_segs = sorted(
+                    [(seg.start, seg.end, label)
+                     for label in currAnnotation.labels()
+                     if label is not None and str(label).strip() != ""
+                     for seg in currAnnotation.subset([label]).itersegments()],
+                    key=lambda x: x[0]
+                )
+                prev_end = 0.0
+                for start, end, label in all_segs:
+                    if start > prev_end + 0.1:
+                        noVoice[Segment(prev_end, start)] = 'silence'
+                    oneVoice[Segment(start, end)] = label
+                    prev_end = max(prev_end, end)
+                if currTotalTime > prev_end + 0.1:
+                    noVoice[Segment(prev_end, currTotalTime)] = 'silence' 
         sumNoVoice    = su.sumTimes(noVoice)
         sumOneVoice   = su.sumTimes(oneVoice)
         sumMultiVoice = su.sumTimes(multiVoice)
-        print(f"DEBUG analyze({inFileName}): noVoice={sumNoVoice:.1f}s oneVoice={sumOneVoice:.1f}s multiVoice={sumMultiVoice:.1f}s total={currTotalTime:.1f}s")
+
 
         # df3
         df3 = utils.build_df3(noVoice, oneVoice, multiVoice)