Fix chord detection - reduce noise, enforce minimum duration

- Increase hop_length from 512 to 2048 (~93ms resolution vs 23ms)
- Heavy median filter smoothing (21 frames = ~2 seconds)
- Higher confidence threshold (0.35 vs 0.12)
- Enforce 1 second minimum chord duration
- Two-pass detection: raw detection then consolidation

This reduces output from 993 noisy changes to ~50-100 musical chord changes.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

modules/chords.py

@@ -270,29 +270,72 @@ def match_chord_template_with_confidence(chroma_frame, templates, focus='harmony
     return best_chord, best_score
 
 
-def extract_chords(audio_path, min_duration=0.5):
-    """Fallback single-file"""
+def extract_chords(audio_path, min_duration=1.0):
+    """
+    Extract chords from audio file with musical timing.
+
+    Args:
+        audio_path: Path to audio file
+        min_duration: Minimum chord duration in seconds (default 1.0s = reasonable for most music)
+
+    Returns:
+        List of (timestamp, chord_name) tuples
+    """
     if not LIBROSA_AVAILABLE:
         return []
-    
+
     try:
+        # Load audio
         y, sr = librosa.load(audio_path, sr=22050, duration=None)
-        chroma = librosa.feature.chroma_cqt(y=y, sr=sr, hop_length=512)
-        chroma = scipy.ndimage.median_filter(chroma, size=(1, 9))
-        
+
+        # Use larger hop for coarser time resolution (fewer false changes)
+        # hop_length=2048 at 22050Hz = ~93ms per frame
+        hop_length = 2048
+
+        # Extract chroma features
+        chroma = librosa.feature.chroma_cqt(y=y, sr=sr, hop_length=hop_length)
+
+        # Heavy smoothing to reduce noise - median filter across 21 frames (~2 seconds)
+        chroma = scipy.ndimage.median_filter(chroma, size=(1, 21))
+
         templates = create_chord_templates()
-        chords = []
-        last_chord = None
-        
+
+        # First pass: detect all chord candidates
+        raw_chords = []
         for i in range(chroma.shape[1]):
             frame = chroma[:, i]
-            time = librosa.frames_to_time(i, sr=sr, hop_length=512)
+            time = librosa.frames_to_time(i, sr=sr, hop_length=hop_length)
             chord, conf = match_chord_template_with_confidence(frame, templates, 'harmony')
-            
-            if chord != last_chord and conf > 0.12:
-                chords.append((float(time), chord))
-                last_chord = chord
-        
-        return chords
-    except:
+            raw_chords.append((float(time), chord, float(conf)))
+
+        # Second pass: consolidate with minimum duration and higher confidence
+        consolidated = []
+        current_chord = None
+        current_start = 0.0
+        confidence_threshold = 0.35  # Higher threshold for cleaner detection
+
+        for time, chord, conf in raw_chords:
+            if conf >= confidence_threshold:
+                if chord != current_chord:
+                    # Check if previous chord lasted long enough
+                    if current_chord is not None:
+                        duration = time - current_start
+                        if duration >= min_duration:
+                            consolidated.append((current_start, current_chord))
+
+                    # Start new chord
+                    current_chord = chord
+                    current_start = time
+
+        # Don't forget the last chord
+        if current_chord is not None and raw_chords:
+            final_time = raw_chords[-1][0]
+            duration = final_time - current_start
+            if duration >= min_duration:
+                consolidated.append((current_start, current_chord))
+
+        return consolidated
+
+    except Exception as e:
+        print(f"   [WARN] Chord extraction error: {e}")
         return []