Repository navigation
Fix/settings scroll align - #7
Merged
Merged
Conversation
qwen3-asr.cpp v0.1.9 forces the spoken language from the prompt when -l is given, the same way the qwen-asr Python package does. Auto-detection is unreliable on short clips (Danish came out as Swedish or Chinese). "auto" keeps detection; older engine builds only used -l for Korean word splitting, so the flag is harmless there. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
v0.1.9 (niksedk/qwen3-asr.cpp#11, SubtitleEdit#12) fixes the runtime-generated mel filterbank: it used the HTK mel scale sampled at integer bins while the Whisper feature extractor uses Slaney-scale filters, so the log-mel input was off by up to 1.6 and the model mis-detected languages on short clips. On 600 held-out Danish clips the stock 0.6B GGUF goes from 49.5 to 43.1 WER, and a Danish fine-tune from 26.3 to 18.8 (transformers: 18.6). The release also pads the trailing conv chunk like the reference and forces the language from the prompt when -l is given (the forced aligner shares the fixed filterbank, so timestamps improve as well). New archive and executable hashes prepended for all seven platforms so existing installs are offered the update. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Adds qwen3-asr-0.6b-da-q8_0.gguf (https://huggingface.co/niksedk/qwen3-asr-0.6b-danish), a LoRA fine-tune of Qwen3-ASR-0.6B on 399 h of CoRal, NST and FTSpeech, merged and quantised to q8_0. On 600 held-out Danish clips through qwen3-asr-cli v0.1.9 it scores 18.8 WER versus 43.1 for the stock 0.6B model. Needs the v0.1.9 engine (SubtitleEdit#16244); older engines run it but lose ~7 WER points to the mel filterbank bug. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…asr-pass-language Qwen3 ASR: pass the selected language to qwen3-asr-cli
…asr-cpp-v019 Qwen3 ASR: bump qwen3-asr.cpp to v0.1.9 (mel filterbank fix)
…asr-danish-model Qwen3 ASR: add Danish fine-tuned 0.6B model
…umeric-field-border Fix compact numeric field borders in EditorChrome
…titleEdit#16031) PR SubtitleEdit#16207 taught the fix engine that a slash standing alone is a slanted italic "I". The fix engine runs on every OCR engine's output, so the title card "PILOTEN / OMGEKOMEN" read by llama.cpp (GLM-OCR) came out as "PILOTEN I OMGEKOMEN" in 5.4.0-beta2. - New IOcrFixEngine.FixImageMatcherMisreads flag, set by the OCR window for nOCR and binary image compare only (next to FillUnknownCharacters). Text engines (Tesseract, Paddle, Ollama, llama.cpp) never fit a glyph to "/" or "#", so the slash and hash misread rules stay off for them. - Even for image matchers a lone slash next to a number ("1 / 2") or between two all-caps words ("PILOTEN / OMGEKOMEN", "AC / DC") is kept as a separator. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
nOCR reads the sans-serif capital I as "l" ("PlLOTEN / OMGEKOMEN"), and
the all-caps l->I rule only runs after the slash rule, so the guard saw a
lowercase letter and still turned the slash into "I". Verified on nl.sup:
the title card now keeps its slash, no other line changes.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…tchers (SubtitleEdit#16031) nOCR fits the sans-serif capital I and lowercase l to the same bar, so on the Dutch Blu-ray names came out as "CowsIip", "HeinkeIs", "FIight" and all-caps words as "JUNl", "lFF", "WAARSCHlJNLlJK". No dictionary knows the names, so for a word that is unknown anyway the letters around the bar decide (TryFixBarMisreadByCase): an I after a lowercase letter, or between the capital and the first lowercase letter, is an l; an l among capitals (two, or one with the bar in front) is an I. Mc/Mac, "Il" and words with another lowercase letter are left alone. Only the English replace list had the all-caps rule; this covers every language. Two italic spacing artifacts are repaired on the line: "20. 000" -> "20.000" (one-to-three digit group, never after a year) and the space before a closing quote after a sentence end ("voorbij. '"). All four run only for image-matcher output (FixImageMatcherMisreads). nl.sup: 45 -> 18 lines differing from the GLM reference, 9 of them GLM's own errors. Before/after on 19 other discs (English, Danish, Polish): every changed line is a repair (RIDDICK, DIDN'T, FBI-enhed, Statler, mathlete, Allenwood, trailing quotes). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…j/v/w
An overhanging f, r or t leans over the j, v or w next to it, so nOCR's
space detector swallows the word space: "Hoorje", "kentje", "achterje",
"Vijfvan", "ofwe". For a word the dictionary does not know, split at that
boundary when both halves are words and the right half is a function word
of at most three letters - without the cap "Getverdikkeme" (not in the
dictionary) became "Get verdikkeme". Image-matcher output only.
nl.sup: 18 -> 13 lines differing from the GLM reference; the other discs
gain one repair ("Alt vel") and lose nothing.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
v0.1.10 (niksedk/qwen3-asr.cpp#13) adds beam search. Subtitle Edit now passes --beam-size 3, which took a Danish fine-tune from 25.2 to 23.4 WER on 300 held-out clips for about twice the decode time (beam 5: 23.1 at four times). Older engine builds reject unknown arguments, so the flag is only passed when the installed qwen3-asr-cli advertises it in --help; the probe runs once per executable. New archive and executable hashes prepended for all seven platforms. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…asr-cpp-v0110-beam Qwen3 ASR: bump qwen3-asr.cpp to v0.1.10 and use beam search (3)
…-issue-analysis-eaf387 OCR fix engine: image-matcher I/l and spacing repairs, lone slash kept (SubtitleEdit#16031)
The model prefixes its output with "language <Name><asr_text>". qwen3-asr.cpp v0.1.9+ strips the marker, but older engine builds leak "<asr_text>" into the aligned word list for GGUFs that store the token as an ordinary token (the Danish fine-tune does), so it appeared at the start of subtitle lines. Strip it in the JSON parser regardless of engine version and drop empty tokens. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…asr-strip-marker Qwen3 ASR: strip the <asr_text> marker from engine output
…izable lists (SubtitleEdit#16251) - The saved styles list gets its category picker back, with new/rename/ delete category beside it. Saved styles are kept per series in categories; with hundreds of them a flat list with a category column was unusable, and the filter and commands were still in the view model. - Both style lists show the font name and a color sample ("Aa" in the primary color on the outline color, italic when the style is), so the colors of every style are visible without selecting it - as SE 5.3.0 and SE4's advanced styles list showed them. - A splitter between the lists and the preview/editor lets the lists pane grow; its width is remembered (SeAssa.StylesListsWidth, default 480 instead of a fixed 360) and shared by the ASSA and SSA windows. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Updated the Japanese translation for Subtitle Edit 5.4.0 Beta2, in preparation for Beta3.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…leedit-issue-16251-08f77e ASSA/SSA styles: category picker back, font name + color columns, resizable lists (SubtitleEdit#16251)
…ian by their letters Natural Macedonian text was reported as Bulgarian (or Russian): the South Slavic languages share their short words (не, да, за, нас, може, така, беше, ...), so the Bulgarian and Russian word lists won, and the Macedonian list was mostly words from one space film plus OCR misreads with Latin letters. The spelling is unambiguous: ј/љ/њ/џ are written only in Macedonian and Serbian, ќ/ѓ/ѕ only in Macedonian, ћ/ђ only in Serbian, and none of them has ъ/ы/э/ё/й/щ/ь/ю/я. One letter scan per detection now decides "mk" or "sr" in the Russian and Bulgarian blocks and in the letter-statistics fallback (which answered "ru" for every Cyrillic text). Bulgarian/Russian keep winning when their letters outnumber the misread ones in an OCR file. The Macedonian word list is rebuilt from everyday words that Bulgarian and Serbian spell differently. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…nian-language-detection-87cfc8 Language auto-detect: tell Macedonian and Serbian from Bulgarian/Russian by their letters
…ings Zoom in/out tooltips were inverted, "Colora le righe" for the custom error rules toggle, "Elenco elenco", "uan", "delal", "fotgramma", "colroe", "Analisi in corso...", double spaces, Disinstalla, "un tuo modello", Cornice for the Frame background box, and the Sync to audio hint now names the real menu items. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Updated the Japanese translation for Subtitle Edit 5.4.0 Beta2, in preparation for Beta3.
Update Italian.json
"Save as iTunes Timed Text" produced a different header for every source format: an ITT source kept its own styling (e.g. tts:color="#EBEBEB"), a TTML source kept its header, and an SRT got the built-in one with white styles, a <body> without a default region and region="bottom" on every line. There was no way to get the same header for every file (SubtitleEdit#16257). File > Properties for iTunes Timed Text now has: - Presets: a named copy of the whole setup (combo + save/delete like batch convert), stored in Settings.Formats.TimedTextItunesPresets. - Header source: "Current file" (as before) or "Template file" - an .itt/.xml whose head (styles, regions, metadata) and body attributes are used for every saved file; the window lists the template's styles and regions and shows a summary or why the file is rejected. The UI reads the file and hands libse the text (TimedTextItunesHeaderTemplate). - Default color: the color of the built-in header's styles; in current file mode OK recolors the styles that carried the old default color. - Default style/region are persisted and put on the built-in <body>. libse ItunesTimedText.ToText: a template header replaces the source header (frame rate/drop mode follow the current frame rate, the language setting wins when set, unknown paragraph styles are dropped), and a line only names its region when it differs from the body's - with <body region="bottom"> only top-aligned lines get region="top". Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…leedit-discussion-16257-7f3414 iTunes Timed Text: export presets with a template-file header source
A fixed-palette dark theme (polar night background, snow storm text, frost blue accents). It shares the dark Fluent variant, icons and syntax colors with Dark, so IsDarkThemeEnabled covers both, and the dark background and foreground lookups return Nord's colors instead of the user's dark-mode settings. The editor palette and plugin theme colors follow it too. Switching between Dark and Nord keeps the dark variant, so the Fluent region color never reached the windows: dark themes now set the window background through a style, and Settings OK refreshes the grid's alternating row color, which has its own Nord shade. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…or Cyrillic, Baltic, Afrikaans/Dutch and Arabic-script A probe with natural dialogue and narration in all 46 supported languages missed 31 of 230 clips, almost all of them two to five lines long: a lone shared word handed the clip to another language, or no word matched and the letter fallback answered from its four Latin alphabets (Catalan as Czech, Romanian as French, Polish as Spanish, Lithuanian as Latvian, Albanian as French, Slovene as Serbian, Ukrainian as Russian, Afrikaans as Dutch, ...). Word lists: everyday words added for English, Danish, Norwegian, Swedish, Finnish, Estonian, Latvian, Catalan, Romanian, Polish, Czech, Slovak, Slovene, Croatian, Serbian (both scripts), Macedonian, Ukrainian, Portuguese, Afrikaans, Dutch, German, Italian, Hungarian, Greek, Albanian, Farsi, Hebrew, Indonesian, Vietnamese and Tagalog, preferring the spellings the neighbouring languages do not share; Lithuanian and Albanian rebuilt (the old lists were names and lines from one film). Words another language also uses (Danish "sig", Dutch "is", Italian "li", Portuguese "nem", ...) kept out. Letter rules, each a single pass and only when the blocks need them: - Cyrillic: і/ї/є/ґ Ukrainian, ы/э/ё Russian, ъ Bulgarian, on top of the Macedonian/Serbian rule; the letter fallback refines its "ru" answer. - Baltic: ą/ę/ė/į/ų Lithuanian vs ā/ē/ī/ģ/ķ/ļ/ņ/ō Latvian. - Afrikaans/Dutch: y vs z and ij. - Arabic script: پ/چ/ژ/گ/ی/ک Persian, ٹ/ڈ/ڑ/ں/ے/ھ/ہ Urdu, ي/ك/ة/ذ/ض/ظ Arabic. - Letter fallback: languages with letters of their own (pl, cs, sk, hu, ro, tr, lt, lv, is, vi, pt, es, sq) are answered before the four-alphabet score. Decision blocks: Czech and Slovak get blocks of their own and score with their own counts; the Danish/Norwegian "only" lists break ties instead of vetoing both; Greek in the ANSI code-page check must beat the Cyrillic count; the right-to-left letter set is letters only (phrases added spaces). LanguageAutoDetectNaturalTextTest: 460 cases from four sample sets, all languages, at one to ten lines and everything together. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
- Remove unused private methods/fields/constants flagged by IDE0051/IDE0052 (write-only fields, dead helpers, unused libvlc delegates) - Remove public/internal methods in ui and libuilogic with no references (UiUtil helpers, FfmpegGenerator, WhisperHelper, etc.); drop the now empty TimeSpanFormatter and SplitHelper files - Drop constructor-injected services that were never used and the IDoSpell parameter of SpellCheckWordLists; update tests accordingly - Remove usings that became unused in the touched files libse public API and the netstandard-only CharLookup paths are untouched. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
New syntax coloring setting "Color if gap is in range" (off by default) with a min/max gap range (defaults 3-12 frames / 120-480 ms, inclusive). It marks mid-sized gaps for the broadcast rule that a gap is either closed (about 2 frames) or at least about 12 frames. Follows ms/frames time code mode like the minimum gap; in frame mode the gap is rounded to whole frames first. Grid gap cell, screen-reader text, HasErrors and the error list (new "Gap in range" class) all use the same check. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…leanup-71a4b7 Code cleanup: remove unused methods, fields and dependencies
Removes using directives flagged by IDE0005 in ui, libuilogic, seconv and the test projects. libse is left alone: it multi-targets netstandard2.1 and net10.0, so a using unused in one target can still be needed by the other. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…usings # Conflicts: # src/libuilogic/AudioToText/WhisperHelper.cs # src/libuilogic/Translate/Formatting.cs # src/ui/Features/Shared/PickRuleProfile/PickRuleProfileViewModel.cs # src/ui/Features/Video/TextToSpeech/TextToSpeechViewModel.cs # tests/UI/Features/SpellCheck/HyphenatedNameSpellCheckTests.cs # tests/UI/Features/SpellCheck/SpanishSeWordListTests.cs # tests/UI/Features/SpellCheck/UserDictionaryNormalizationTests.cs # tests/UI/Features/Sync/PointSync/PointSyncWindowTests.cs # tests/UI/Features/Video/EmbeddedSubtitlesEditTests.cs # tests/UI/Features/Video/TextToSpeechWindowTests.cs
…gap-in-range Color the grid gap when it falls inside a user-defined range
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…-unused-usings Remove unused usings outside libse
Japanese translation for Subtitle Edit 5.4.0 Beta3
Update Italian.json
Update Japanese.json for Subtitle Edit 5.4.0 Beta4
Grid copy/cut and waveform copy wrote the video-relative row times to the clipboard, while the grid shows and a save writes them with the video offset added. Copy now adds the offset and the subtitle-format paste takes it off again, so copy + paste inside SE keeps the timing. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Remember the last copied text and its video offset; when the clipboard still holds that copy, paste subtracts that offset instead of the current one, so copy, change offset, paste keeps the original times. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…SubtitleEdit#16095) New Settings > Tools switch "Split at video and text box position: cursor at the end keeps all text in the first line (second line empty)". On, the "split at video and text box position" commands (and their "... and play" variants) with the caret at the end keep all the text in the first half and leave the second empty, like SE 4. Off (default) keeps the SubtitleEdit#14962 fallback to the line break. Plain "split at cursor" and the original text box are not affected. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…offset-grid-copy-73e306 Include video offset when copying from the grid
…leedit-issue-16095-0c0fe3 Setting: split at video and text box position keeps text at caret end (SubtitleEdit#16095)
Hunspell knows rare words like Dutch "weetje", "watje", "wasje" and "moetje", so an OCR'd "weet je" with the space lost was never flagged by spell check, and the OCR fix rules that only fire on unknown words left it alone. Options > Word lists now has a "Removed from dictionary" list: those words count as misspelled whatever Hunspell says, are not offered as suggestions, and the OCR fix pronoun-split rule can fix them. Stored in a <removed> section of <lang>_user.xml (also read from the shipped _se.xml lists; the user's own file wins). Adding a word to the user dictionary takes it off the removed list and the other way round. Writing the user word list no longer drops the rest of the file. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Shows the current speed (e.g. 1.0x) in the video controls; a click opens a menu with the speeds. Hidden by default, configurable in Video controls. Goes through SelectedSpeed, so the waveform toolbar speed follows, and the full-screen player now gets the speed on open and hands it back on close. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…heck-rejected-words Spell check: remove words from the dictionary (SubtitleEdit#16031)
…ck-speed-video-player-900687 Video player: optional playback speed button (SubtitleEdit#16306)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…-beta4 v5.4.0-beta4
Update Korean Translation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9c80b58b8a
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| internal static byte[] ReadPayload(System.IO.Stream fs, ulong length) | ||
| { | ||
| var bytesLeft = fs.Length > fs.Position ? (ulong)(fs.Length - fs.Position) : 0UL; | ||
| var buffer = new byte[Math.Min(length, bytesLeft + 8)]; |
There was a problem hiding this comment.
When a truncated MP4 ends immediately after a child-box header, bytesLeft is zero and this returns only eight bytes even if the declared payload is much larger. Callers then immediately read beyond that buffer—for example, Stbl reads GetUInt(8) for stsz, while Mdia reads the handler type at offset 8—so opening such a malformed file now throws ArgumentOutOfRangeException instead of safely rejecting it. Either pad to each caller's required header size as before or make callers validate the returned length before indexing.
Useful? React with 👍 / 👎.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.