#!/usr/bin/env python3 """Generate V2 subtitles with correct text + known voiceover segment timing""" import json, os, re BASE = r'E:\集群文件夹\factory_os\short_video_real_data_pipeline\phase4b_90s_formal_sample\v22_kimi_claude_loop\full_story_douyin_video' SUB_DIR = os.path.join(BASE, '04_subtitles', 'v2') os.makedirs(SUB_DIR, exist_ok=True) # Correct script text (9 segments, matching the voiceover segments) SCRIPT_SEGMENTS = [ "我让 Claude Code 自己做视频。录屏失败、AI 跑偏、从 69 爬到 92。这不是特效，这是真实工作流。", "Windows 后台抓不到桌面画面。试了五种录屏方案，全是空文件。录屏路线，彻底失败。", "更糟的是，Claude Code 跑去研究亚马逊选品。ChatGPT 看不下去：放弃录屏，让它自己画界面。", "于是用 Python 画终端界面、工具调用卡片。四个 AI 代理并行跑，进度条在涨，代码在敲。每条命令都来自真实执行日志。", "每轮渲染完 Kimi 审核打分。v1 69 分，改字体 v2 72，加动画 v3 跳到 81。修字幕 v4 85，砍时长 v6 89。V9 冲到 92。", "但下一轮有个硬编码没改。画面四个代理，底部只显示三个。92 直接砸到 76。", "修了六项回到 88，但回不到 92 了。一个硬编码，一夜回到解放前。最终决定：回滚 V9，冻结。", "V9 上线审核门户，HTTP 200 通过。11 轮迭代，1 次崩溃，从 69 到 92。AI 自己审核自己，自己修复自己，自己部署自己。", "录屏失败、跑偏、崩过、又爬回来。执行、审核、修复、回归、交付。这才是真正的 AI 工作流。", ] # Original segment durations before atempo SEG_DURATIONS = [10.1, 8.6, 9.0, 12.3, 14.8, 8.4, 10.1, 13.3, 10.6] ATEMPO = 1.05 # Calculate segment timing (proportional to atempo) total_orig = sum(SEG_DURATIONS) # 97.2 total_final = total_orig / ATEMPO # ~92.57 seg_timing = [] current_time = 0.0 for dur in SEG_DURATIONS: final_dur = dur / ATEMPO seg_timing.append((current_time, current_time + final_dur)) current_time += final_dur print("Segment timing (seconds):") for i, (st, et) in enumerate(seg_timing): print(f" P{i+1}: {st:.2f}s - {et:.2f}s ({et-st:.2f}s)") # Keywords to highlight with colors KEYWORD_MAP = { "录屏失败": "#FF4444", "录屏": "#FF4444", "AI 跑偏": "#FF8800", "跑偏": "#FF8800", "真实执行日志": "#00DDFF", "真实日志": "#00DDFF", "Kimi": "#00AAFF", "69": "#FFD700", "72": "#88FF88", "81": "#44FF44", "85": "#44FFDD", "89": "#44DDFF", "92 砸到 76": "#FF2222", "92": "#FFD700", "76": "#FF2222", "88": "#FFDD44", "HTTP 200": "#44FF44", "200": "#44FF44", "回滚": "#44FF44", "冻结": "#44FFDD", "稳定交付": "#44FF44", "regression": "#FF2222", "自动修复": "#44DDFF", } def get_kw_color(text): for kw, color in sorted(KEYWORD_MAP.items(), key=lambda x: -len(x[0])): if kw in text: return color, kw return None, None def format_time(seconds): h = int(seconds // 3600) m = int((seconds % 3600) // 60) s = seconds % 60 return f"{h:02d}:{m:02d}:{s:06.3f}" def srt_time(sec): return format_time(sec).replace('.', ',') def ass_time(sec): return format_time(sec).replace(',', '.') # Generate subtitles all_subs = [] sub_idx = 1 for seg_idx, (text, (seg_start, seg_end)) in enumerate(zip(SCRIPT_SEGMENTS, seg_timing)): seg_dur = seg_end - seg_start # Split into sentences sentences_raw = re.split(r'(?<=[。！？])', text) sentences = [s.strip() for s in sentences_raw if s.strip()] # Merge short sentences into groups of max 13 chars per line line_groups = [] current_lines = [] current_chars = 0 for sent in sentences: sent_clean = sent.replace(' ', '') sent_chars = len(sent_clean) # If this sentence alone would exceed 13 chars, it becomes its own line if sent_chars > 13: if current_lines: line_groups.append(list(current_lines)) current_lines = [] current_chars = 0 current_lines.append(sent) current_chars = sent_chars line_groups.append(list(current_lines)) current_lines = [] current_chars = 0 continue if current_chars + sent_chars <= 13: current_lines.append(sent) current_chars += sent_chars # Max 2 lines per group if len(current_lines) >= 2: line_groups.append(list(current_lines)) current_lines = [] current_chars = 0 else: if current_lines: line_groups.append(list(current_lines)) current_lines = [sent] current_chars = sent_chars if current_lines: line_groups.append(list(current_lines)) # If no groups (empty), create one if not line_groups: line_groups.append([text]) # Distribute timing across groups num_groups = len(line_groups) group_chars = [sum(len(l.replace(' ', '')) for l in g) for g in line_groups] total_chars = sum(group_chars) g_start = seg_start for gi, (group, gc) in enumerate(zip(line_groups, group_chars)): g_dur = (gc / total_chars) * seg_dur g_end = g_start + g_dur if g_end - g_start < 0.6: g_end = g_start + 0.6 if gi == num_groups - 1: g_end = seg_end all_subs.append({ "index": sub_idx, "start": round(g_start, 3), "end": round(g_end, 3), "lines": group }) sub_idx += 1 g_start = g_end # --- Generate SRT --- srt_lines = [] for sub in all_subs: srt_lines.append(str(sub['index'])) srt_lines.append(f"{srt_time(sub['start'])} --> {srt_time(sub['end'])}") for line in sub['lines']: srt_lines.append(line) srt_lines.append("") srt_path = os.path.join(SUB_DIR, 'captions_final_v2.srt') with open(srt_path, 'w', encoding='utf-8') as f: f.write('\n'.join(srt_lines)) print(f"\nSRT saved: {srt_path} ({len(all_subs)} subtitles)") print(f" Total duration: {all_subs[-1]['end']:.1f}s") # --- Generate ASS with keyword highlighting --- ass_header = """[Script Info] ScriptType: v4.00+ PlayResX: 1080 PlayResY: 1920 ScaledBorderAndShadow: yes [V4+ Styles] Format: Name, Fontname, Fontsize, PrimaryColour, SecondaryColour, OutlineColour, BackColour, Bold, Italic, Underline, StrikeOut, ScaleX, ScaleY, Spacing, Angle, BorderStyle, Outline, Shadow, Alignment, MarginL, MarginR, MarginV, Encoding Style: Default,Microsoft YaHei,28,&H00FFFFFF,&H0000FF00,&H00000000,&H80000000,1,0,0,0,100,100,0,0,1,2,0,2,50,50,80,134 [Events] Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text """ def apply_highlighting(text): """Apply ASS color tags for keywords""" sorted_kws = sorted(KEYWORD_MAP.items(), key=lambda x: -len(x[0])) result = [] remaining = text while remaining: best_pos = len(remaining) best_kw = None best_color = None for kw, color in sorted_kws: pos = remaining.find(kw) if pos != -1 and pos < best_pos: best_pos = pos best_kw = kw best_color = color if best_kw is not None: if best_pos > 0: result.append(remaining[:best_pos]) # Convert #RRGGBB to BBGGRR for ASS bgr = best_color[5:7] + best_color[3:5] + best_color[1:3] result.append(f"{{\\c&H{bgr}&\\b1}}{best_kw}{{\\c&HFFFFFF&\\b0}}") remaining = remaining[best_pos + len(best_kw):] else: result.append(remaining) break return "".join(result) ass_lines = [ass_header] for sub in all_subs: start_str = ass_time(sub['start']) end_str = ass_time(sub['end']) hl_lines = [apply_highlighting(l) for l in sub['lines']] ass_text = "\\N".join(hl_lines) ass_lines.append(f"Dialogue: 0,{start_str},{end_str},Default,,0,0,0,,{ass_text}") ass_path = os.path.join(SUB_DIR, 'captions_final_v2.ass') with open(ass_path, 'w', encoding='utf-8') as f: f.write('\n'.join(ass_lines)) print(f"ASS saved: {ass_path}") # --- Generate word-level JSON --- words_json = { "total_duration": round(total_final, 1), "model": "script+segment-timing", "segments": [] } for seg_idx, (text, (seg_start, seg_end)) in enumerate(zip(SCRIPT_SEGMENTS, seg_timing)): clean = text.replace(' ', '') cc = len(clean) seg_words = [] if cc > 0: char_dur = (seg_end - seg_start) / cc for ci, ch in enumerate(clean): w_start = seg_start + ci * char_dur w_end = seg_start + (ci + 1) * char_dur # Check if this character is part of a keyword is_hl = False for kw in KEYWORD_MAP: if kw.startswith(ch) or ch in kw: # Approximate: check if this position has a keyword kw_pos = text.replace(' ', '').find(kw.replace(' ', '')) if kw_pos != -1 and kw_pos <= ci < kw_pos + len(kw.replace(' ', '')): is_hl = True break seg_words.append({ "word": ch, "start": round(w_start, 2), "end": round(w_end, 2), "is_highlight": is_hl }) words_json["segments"].append({ "start": round(seg_start, 2), "end": round(seg_end, 2), "text": text, "words": seg_words }) words_path = os.path.join(SUB_DIR, 'captions_words_v2.json') with open(words_path, 'w', encoding='utf-8') as f: json.dump(words_json, f, ensure_ascii=False, indent=2) print(f"Word JSON saved: {words_path}") print(f"\n{'='*50}") print(f"V2 SUBTITLE GENERATION COMPLETE") print(f"{'='*50}") print(f"Total duration: {total_final:.1f}s") print(f"Subtitle groups: {len(all_subs)}") print(f"Script segments: {len(SCRIPT_SEGMENTS)}") for sub in all_subs: print(f" #{sub['index']:2d} [{sub['start']:6.2f}-{sub['end']:6.2f}] {' / '.join(sub['lines'][:2])}")