32baf9cd45Auto-resume: cleanup stuck jobs on container startup + GEMINI_API_KEY envSebastjan Artič2026-04-29 08:43:31 +0000
ec71c54570Upgrade to Sonnet 4.6 + add Gemini 3.1 Pro supportSebastjan Artič2026-04-29 08:26:27 +0000
9faa224885Upgrade Claude model: Haiku 4.5 → Sonnet 4.6 for better Slavic language transcript correctionSebastjan Artič2026-04-29 08:22:10 +0000
69fb2f5ce8Upgrade default Whisper model: small/medium → large-v3 for much better Slovenian/Slavic transcription accuracySebastjan Artič2026-04-29 08:20:18 +0000
4bc5ac6756Major: Claude post-processing of Whisper transcriptSebastjan Artič2026-04-29 08:13:33 +0000
4e123bdabcUI: hide lang/model dropdowns — both are fully automatic now (3-sample lang detection + medium default model)Sebastjan Artič2026-04-29 08:03:22 +0000
c870d80726Fix: extend clip if ends mid-vocal (no chorus cut-off), DejaVu Sans font (supports SLO/HR/BS chars), auto-upgrade to medium Whisper model for Slavic languagesSebastjan Artič2026-04-29 07:35:00 +0000
5d5e169f9dDisable Whisper VAD filter — was dropping vocal segments in songs creating gaps in subtitlesSebastjan Artič2026-04-29 07:07:29 +0000
a04811bdc9Add Claude LLM analysis: sends full transcript to Claude API for true song structure understanding (refrain detection across all repetitions, not just local heuristic)Sebastjan Artič2026-04-29 06:55:41 +0000
e072eec362Fix: handle Whisper transcribe failure for instrumental-only audio (fallback to empty transcript)Sebastjan Artič2026-04-29 06:33:52 +0000
33a138af9eFix: force native Python bool/float for JSON serialization (numpy types)Sebastjan Artič2026-04-29 06:23:41 +0000
81edd24ca3Subtitles: smaller font 56px (was 84), higher position MarginV=400, side margins 80px for safe zoneSebastjan Artič2026-04-29 06:09:26 +0000
ba787744a6Subtitles: cap chunk duration at 2.5s, split long lines into multiple time slices for faster reels pacingSebastjan Artič2026-04-29 05:59:36 +0000
e001387a89Subtitles: convert SRT to ASS directly with PlayResY=1920 for predictable scaling instead of unreliable force_styleSebastjan Artič2026-04-28 18:09:53 +0000
28d933c916Subtitles: UPPERCASE + position lower (MarginV=320 for 1080x1920) + bigger fontSebastjan Artič2026-04-28 17:40:48 +0000
d36893bf2dFIX CRITICAL: reload job dict after find_chorus update so reframe gets new start/duration valuesSebastjan Artič2026-04-28 17:33:11 +0000
8eaef029e2Find chorus: weight repetitive short phrases (like 'Ohne dich x5') as strong chorus signalSebastjan Artič2026-04-28 16:57:45 +0000
c17578521aFix find_chorus: RMS energy parser was broken (no pts_time available), now syntheses timestamps; energy weight x10 (refren je glasnejši)Sebastjan Artič2026-04-28 16:55:51 +0000
64e8854ceaTrack mode: more sensitive face detection + longer smoothing windowSebastjan Artič2026-04-28 16:45:13 +0000