feat: align_segments() — map Whisper timestamps to pyannote speakers

alignment.py

@@ -0,0 +1,32 @@
+def align_segments(
+    whisper_segs: list[dict],
+    speaker_segs: list[tuple[float, float, str]],
+) -> list[tuple[str, str]]:
+    """Assign each Whisper segment to the speaker with the greatest time overlap.
+    Consecutive segments from the same speaker are merged into one block."""
+    result: list[tuple[str, str]] = []
+    for seg in whisper_segs:
+        speaker = _best_speaker(seg["start"], seg["end"], speaker_segs)
+        text = seg["text"].strip()
+        if not text:
+            continue
+        if result and result[-1][0] == speaker:
+            result[-1] = (speaker, result[-1][1] + " " + text)
+        else:
+            result.append((speaker, text))
+    return result
+
+
+def _best_speaker(
+    start: float,
+    end: float,
+    speaker_segs: list[tuple[float, float, str]],
+) -> str:
+    best_label = "SPEAKER_00"
+    best_overlap = 0.0
+    for s_start, s_end, label in speaker_segs:
+        overlap = max(0.0, min(end, s_end) - max(start, s_start))
+        if overlap > best_overlap:
+            best_overlap = overlap
+            best_label = label
+    return best_label

tests/test_alignment.py

@@ -0,0 +1,33 @@
+def test_align_assigns_speaker_by_overlap():
+    from alignment import align_segments
+    whisper = [
+        {"start": 0.0, "end": 2.0, "text": "Hallo"},
+        {"start": 2.1, "end": 4.0, "text": "Wie geht es"},
+    ]
+    speakers = [
+        (0.0, 2.5, "SPEAKER_00"),
+        (2.5, 5.0, "SPEAKER_01"),
+    ]
+    result = align_segments(whisper, speakers)
+    assert result[0] == ("SPEAKER_00", "Hallo")
+    assert result[1] == ("SPEAKER_01", "Wie geht es")
+
+
+def test_align_merges_consecutive_same_speaker():
+    from alignment import align_segments
+    whisper = [
+        {"start": 0.0, "end": 1.0, "text": "Hallo"},
+        {"start": 1.1, "end": 2.0, "text": "Welt"},
+    ]
+    speakers = [(0.0, 3.0, "SPEAKER_00")]
+    result = align_segments(whisper, speakers)
+    assert len(result) == 1
+    assert result[0] == ("SPEAKER_00", "Hallo Welt")
+
+
+def test_align_fallback_when_no_speaker_overlap():
+    from alignment import align_segments
+    whisper = [{"start": 0.0, "end": 1.0, "text": "Hallo"}]
+    speakers = []
+    result = align_segments(whisper, speakers)
+    assert result[0][0] == "SPEAKER_00"