transcribe: classify reference lines as matched/dropped, flag unmatched singing

src/songtrace/stages/transcribe.py

@@ -20,10 +20,56 @@ def _tokens(text: str) -> list[str]:
     return [t for w in text.split() if (t := alignment_token(w))]
 
 
+def match_lyrics(segments, ref: RefLyrics):
+    heard, seg_of = [], []
+    for si, s in enumerate(segments):
+        for t in _tokens(s.text):
+            heard.append(t)
+            seg_of.append(si)
+    pairs = [(t, li) for w, li in zip(ref.words, ref.word_line_index)
+             if (t := alignment_token(w))]
+    expected = [t for t, _ in pairs]
+    line_of = [li for _, li in pairs]
+    matcher = difflib.SequenceMatcher(a=heard, b=expected)
+    return matcher, seg_of, line_of
+
+
 def lyrics_similarity(segments, ref: RefLyrics) -> float:
-    heard = [t for s in segments for t in _tokens(s.text)]
-    expected = [t for w in ref.words if (t := alignment_token(w))]
-    return difflib.SequenceMatcher(a=heard, b=expected).ratio()
+    matcher, _, _ = match_lyrics(segments, ref)
+    return matcher.ratio()
+
+
+def classify_lyrics_match(segments, ref: RefLyrics, sections
+                          ) -> tuple[list[int], list[tuple[float, float]]]:
+    """Split reference lines into matched vs. dropped, and flag sung spans no
+    reference line explains.
+
+    dropped_lines: reference line indices covered only by 'insert' opcodes
+    (present in the lyrics, never heard). missing_spans: vocal `sections`
+    containing 'delete' opcodes (heard, but matching no reference word).
+    """
+    matcher, seg_of, line_of = match_lyrics(segments, ref)
+    covered = [False] * len(line_of)
+    for tag, _, _, bj1, bj2 in matcher.get_opcodes():
+        if tag in ("equal", "replace"):
+            for j in range(bj1, bj2):
+                covered[j] = True
+    by_line = {}
+    for j, li in enumerate(line_of):
+        by_line.setdefault(li, []).append(covered[j])
+    dropped_lines = sorted(li for li, cov in by_line.items() if not any(cov))
+
+    missing = set()
+    for tag, ai1, ai2, _, _ in matcher.get_opcodes():
+        if tag != "delete":
+            continue
+        si = seg_of[ai1]
+        mid = (segments[si].start + segments[si].end) / 2
+        for a, b in sections:
+            if a <= mid <= b:
+                missing.add((a, b))
+                break
+    return dropped_lines, sorted(missing)
 
 
 def _run(ctx: Ctx) -> None:

tests/test_transcribe.py

@@ -1,6 +1,6 @@
 from songtrace.lyrics import normalize_lyrics
 from songtrace.whisper_client import TranscriptSegment
-from songtrace.stages.transcribe import lyrics_similarity, vocal_sections
+from songtrace.stages.transcribe import classify_lyrics_match, lyrics_similarity, vocal_sections
 
 def seg(s, e, t=""):
     return TranscriptSegment(start=s, end=e, text=t)
@@ -26,3 +26,29 @@ def test_similarity_low_for_wrong_lyrics():
     ref = normalize_lyrics("completely different words here entirely")
     segs = [seg(0, 2, "it's one two three four five six seven eight")]
     assert lyrics_similarity(segs, ref) < 0.4
+
+def test_classify_lyrics_match_flags_extra_reference_line():
+    ref = normalize_lyrics("it's one two three\nfour five six seven eight\nnine ten eleven")
+    segs = [seg(1, 4, "it's one two three four five six seven eight")]
+    sections = vocal_sections(segs)
+    dropped, missing = classify_lyrics_match(segs, ref, sections)
+    assert dropped == [2]
+    assert missing == []
+
+def test_classify_lyrics_match_flags_unmatched_singing():
+    ref = normalize_lyrics("it's one two three\nfour five six seven eight")
+    segs = [seg(1, 4, "it's one two three four five six seven eight"),
+            seg(20, 23, "completely unrelated ad lib bridge")]
+    sections = vocal_sections(segs)
+    dropped, missing = classify_lyrics_match(segs, ref, sections)
+    assert dropped == []
+    assert missing == [(19.0, 24.0)]
+
+def test_classify_lyrics_match_keeps_partially_matched_lines():
+    ref = normalize_lyrics("it's one two three\nfour five six seven eight")
+    # line 1 heard with one word garbled ("five" -> "fife"): still a replace/equal
+    # match, so the line stays matched rather than dropped
+    segs = [seg(1, 4, "it's one two three four fife six seven eight")]
+    sections = vocal_sections(segs)
+    dropped, _ = classify_lyrics_match(segs, ref, sections)
+    assert dropped == []