Fix find_chorus: RMS energy parser was broken (no pts_time available), now syntheses timestamps; energy weight x10 (refren je glasnejši)

scripts/find_chorus.py

@@ -119,33 +119,42 @@ def find_repeated_lines(lines, similarity_threshold=0.5):
 
 def compute_energy(audio_path, window_sec=1.0):
     """
-    Vrni list (timestamp, rms_db) preko FFmpeg ebur128 filter.
+    Vrni list (timestamp, rms_db) preko FFmpeg astats filter.
+    Vsako okno window_sec sekund vrne en RMS sample.
     """
-    # Uporabi ebur128 ali astats za RMS
     cmd = [
         "ffmpeg", "-i", audio_path,
         "-af", f"asetnsamples=n={int(16000 * window_sec)}:p=0,astats=metadata=1:reset={window_sec},"
-               "ametadata=print:key=lavfi.astats.Overall.RMS_level",
+               "ametadata=print:key=lavfi.astats.Overall.RMS_level:file=-",
         "-f", "null", "-",
     ]
     result = subprocess.run(cmd, capture_output=True, text=True)
-    output = result.stderr
+    # ametadata file=- pošilja na stdout
+    output = result.stdout + "\n" + result.stderr
 
     energies = []
     current_pts = None
     for line in output.split("\n"):
         line = line.strip()
-        if line.startswith("frame:"):
-            # frame:N pts:X pts_time:Y
-            m = re.search(r"pts_time:(\S+)", line)
-            if m:
+        # Format A: "frame:N pts:X pts_time:Y"
+        m = re.search(r"pts_time:(\S+)", line)
+        if m:
+            try:
                 current_pts = float(m.group(1))
-        elif line.startswith("lavfi.astats.Overall.RMS_level="):
-            val = line.split("=")[1]
+            except ValueError:
+                pass
+            continue
+        # Format B: lavfi.astats.Overall.RMS_level=-15.123
+        if "RMS_level=" in line:
+            val = line.split("RMS_level=")[-1].strip()
             try:
                 rms = float(val)
-                if current_pts is not None:
-                    energies.append((current_pts, rms))
+                # Če nimamo timestamp-a, sintetiziraj na podlagi vrstnega reda
+                if current_pts is None:
+                    current_pts = len(energies) * window_sec
+                energies.append((current_pts, rms))
+                # Increment za naslednji vzorec, če FFmpeg ne pošilja pts
+                current_pts += window_sec
             except ValueError:
                 pass
 
@@ -210,11 +219,12 @@ def find_chorus(video, lang=None, model_size="small", target_duration=30.0):
                 avg_e = avg_energy_in_range(energies, start, start + target_duration)
                 energy_score = max(0, avg_e - avg_overall)  # koliko nad povprečjem
 
-                # Score: število ponovitev + energy + dolžina vrstice
+                # Score: ponovitve + energija + dolžina vrstice
+                # Refren je navadno glasnejši kot verz — energija je močnejši signal
                 score = (
-                    len(cluster) * 10  # repetition weight
-                    + energy_score * 2  # energy weight
-                    + min(len(rep_text.split()), 10)  # text richness
+                    len(cluster) * 5  # repetition weight (zmanjšano)
+                    + energy_score * 10  # energy weight (povečano — refren je glasnejši)
+                    + min(len(rep_text.split()), 10)
                 )
 
                 candidates.append({
@@ -229,11 +239,11 @@ def find_chorus(video, lang=None, model_size="small", target_duration=30.0):
                     "cluster_id": cluster_idx,
                 })
 
-        # Sort by score, dedupe close candidates
+        # Sort by score, dedupe close candidates (vsaj 20s narazen)
         candidates.sort(key=lambda c: -c["score"])
         deduped = []
         for c in candidates:
-            if all(abs(c["start"] - d["start"]) > 5 for d in deduped):
+            if all(abs(c["start"] - d["start"]) > 20 for d in deduped):
                 deduped.append(c)
             if len(deduped) >= 5:
                 break