From d3b71942d2048ac0324285251cc0c1f08083219f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Sebastjan=20Arti=C4=8D?= <sebastjan@folx.tv>
Date: Wed, 29 Apr 2026 16:52:44 +0000
Subject: [PATCH] Word-level extension: 2-word lookback (not full phrase)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Refinement of previous lookback fix - limit to MAX 2 words back.

Reason: with unlimited lookback, the lookback would chain through
words with gaps < 0.5s and keep walking back into the previous verse.
For Žena ME TEPE: 'verjet.' [76.78] → 'Žena' [76.88] gap is 0.10s,
which means lookback would walk back to verses before chorus.

With 2-word limit:
- Clip at 78.19s → 'me' [78.16] is closest preceding word (gap 0.03s)
- Lookback j=i: 'me' → 'Žena' gap 0.14s → captured (i-1)
- Lookback j=i-1: 'Žena' → 'verjet.' gap 0.10s → would be captured
  but MAX_LOOKBACK_WORDS=2 stops here ✓

Result: anchor = 'Žena' at 76.88s → new_start = 76.73s.
Subtitle: 'ŽENA ME TEPE' (full phrase, no verse leakage).
---
 scripts/analyze.py | 41 ++++++++++++++++++++++++++++++++---------
 1 file changed, 32 insertions(+), 9 deletions(-)

diff --git a/scripts/analyze.py b/scripts/analyze.py
index 4612287..5146e8b 100644
--- a/scripts/analyze.py
+++ b/scripts/analyze.py
@@ -1436,27 +1436,50 @@ def main():
             if all_words:
                 # Najdi "rob" — beseda kjer končanje zelo blizu clip start
                 # ALI clip start je sredi besede (besedo bi odrezali)
+                # ALI prejšnje besede so del istega govora pred clip start
+                
+                # Strategija: poišči besedo PRED clip start, nato razširi nazaj
+                # za **1-2 besedi** (ne celo frazo - to bi zajelo prejšnji verz).
+                # Kombiniraj z amplitude defense (Layer 3) ki dodatno doda buffer.
+                MAX_LOOKBACK_WORDS = 2  # max 2 besedi nazaj
+                
                 for i, w in enumerate(all_words):
                     # Beseda zaobsega clip start (clip reže sredi besede)
                     if w["start"] < current_start < w["end"]:
-                        new_start = max(0, w["start"] - 0.15)
+                        # Razširi nazaj na začetek te besede in največ MAX_LOOKBACK_WORDS predhodnih
+                        anchor_idx = i
+                        for j in range(i, max(0, i - MAX_LOOKBACK_WORDS), -1):
+                            prev = all_words[j - 1]
+                            curr = all_words[j]
+                            gap = curr["start"] - prev["end"]
+                            if gap >= 0.5:
+                                break
+                            anchor_idx = j - 1
+                        new_start = max(0, all_words[anchor_idx]["start"] - 0.15)
+                        captured = " ".join(w2["text"].strip() for w2 in all_words[anchor_idx:i+1])
                         print(f"   🎵 Razširim clip začetek {current_start:.2f}s → {new_start:.2f}s "
-                              f"(clip rezal sredi besede '{w['text'].strip()}')", file=sys.stderr)
+                              f"(clip sredi besede; ujamem '{captured}')", file=sys.stderr)
                         current_start = new_start
                         break
                     # Beseda končana TIK pred clip start (do 0.5s pred)
-                    # IN je naslednja beseda PO/blizu clip start
                     if 0 < (current_start - w["end"]) <= 0.5:
                         # Preveri naslednjo besedo
                         next_w = all_words[i + 1] if i + 1 < len(all_words) else None
                         if next_w and next_w["start"] >= current_start - 0.1:
-                            # Razdalja od te besede do naslednje > 0.3s pomeni mogoče prelom verz/refren
-                            gap_to_next = next_w["start"] - w["end"]
-                            # Razširi nazaj na začetek te besede - 0.15s buffer
-                            new_start = max(0, w["start"] - 0.15)
+                            # Najdi anchor: do MAX_LOOKBACK_WORDS nazaj
+                            anchor_idx = i
+                            for j in range(i, max(0, i - MAX_LOOKBACK_WORDS), -1):
+                                prev = all_words[j - 1]
+                                curr = all_words[j]
+                                gap = curr["start"] - prev["end"]
+                                if gap >= 0.5:
+                                    break
+                                anchor_idx = j - 1
+                            new_start = max(0, all_words[anchor_idx]["start"] - 0.15)
+                            captured = " ".join(w2["text"].strip() for w2 in all_words[anchor_idx:i+1])
                             print(f"   🎵 Razširim clip začetek {current_start:.2f}s → {new_start:.2f}s "
-                                  f"(beseda '{w['text'].strip()}' končana {current_start - w['end']:.2f}s pred clip start, "
-                                  f"morda začne refren; gap do '{next_w['text'].strip()}' = {gap_to_next:.2f}s)", file=sys.stderr)
+                                  f"(beseda '{w['text'].strip()}' tik pred clip start; "
+                                  f"ujamem celo frazo '{captured}')", file=sys.stderr)
                             current_start = new_start
                             break
             else: