v0.6.5

Which file do I download?

file	who it is for	what it does
transcript-normalizer-0.6.5-<system>-lite	anyone who just wants to use it	fetches the platform's captions and fixes domain terms; opens a menu when double-clicked
transcript-normalizer-0.6.5-<system>-full (macOS, Linux)	anyone who also has videos without captions, or audio files	everything lite does, and also transcribes audio on your computer (the first time, it downloads a 1.5 GB speech model). On macOS, a .zip: unzip, then open transcript-normalizer inside the folder. On Linux, a .tar.gz of a folder: extract it, then run transcript-normalizer inside the folder. On Windows, full is the two files below
transcript-normalizer-0.6.5-windows-full-setup.exe	Windows, full: the installer	choose where it goes (for you only, no administrator needed, or for all users); Start menu shortcut; clean uninstall from "Add or remove programs", which keeps your files
transcript-normalizer-0.6.5-windows-full-portable.zip	Windows, full: portable	unzip anywhere, e.g. a USB drive, and open transcript-normalizer.exe inside the folder; everything (your files, settings, the speech model) stays in its data folder; delete the folder to remove it
transcript_normalizer-0.6.5-py3-none-any.whl, .tar.gz	Python developers	the package, for pip install

<system> is windows, macos (Apple silicon) or linux; Linux executables run on glibc 2.35 or newer (Ubuntu 22.04 and later). The executables are not signed, so Windows and macOS will warn the first time; the SHA-256 of every file is at the end of this page.

Qual arquivo baixar?

arquivo	para quem	o que faz
transcript-normalizer-0.6.5-<sistema>-lite	quem só quer usar	pega a legenda da plataforma e corrige os termos da área; abre um menu com dois cliques
transcript-normalizer-0.6.5-<sistema>-full (macOS, Linux)	quem também tem vídeos sem legenda, ou arquivos de áudio	tudo o que o lite faz, e ainda transcreve o áudio no seu computador (na primeira vez, baixa um modelo de fala de 1,5 GB). No macOS, um .zip: descompacte e abra o transcript-normalizer dentro da pasta. No Linux, um .tar.gz de uma pasta: extraia e rode o transcript-normalizer dentro da pasta. No Windows, o full são os dois arquivos abaixo
transcript-normalizer-0.6.5-windows-full-setup.exe	Windows, full: o instalador	você escolhe onde instalar (só para você, sem administrador, ou para todos os usuários); atalho no menu Iniciar; desinstalação limpa em "Adicionar ou remover programas", que mantém os seus arquivos
transcript-normalizer-0.6.5-windows-full-portable.zip	Windows, full: portátil	descompacte em qualquer lugar, por exemplo num pendrive, e abra o transcript-normalizer.exe dentro da pasta; tudo (seus arquivos, configurações, o modelo de fala) fica na pasta data dele; apague a pasta para removê-lo
transcript_normalizer-0.6.5-py3-none-any.whl, .tar.gz	quem programa em Python	o pacote, para pip install

<sistema> é windows, macos (Apple Silicon) ou linux; os executáveis de Linux rodam com glibc 2.35 ou mais nova (Ubuntu 22.04 em diante). Os executáveis não são assinados, então o Windows e o macOS avisam na primeira vez; o SHA-256 de cada arquivo está no fim desta página.

---

Other sites besides YouTube: their sign-in, login walls and missing captions (D-078, D-079); a review with one stop per form where Enter decides nothing (D-072, D-073); the caption in the video's own language, the channel's own first (D-080, D-081); fewer wrong guesses, from a common-word guard and number words matched as digits (D-083, D-084); applied and suggested counted apart (D-082, D-085).

• "applied, please confirm N" now counts only the close guesses that are in normalized.txt; names found by sound, which are never applied until you confirm them, have their own line, "suggested, not applied M", in the summary, the menu and report.txt (D-085, amending D-082).
• Number words from 2 to 20 match their digits (D-084), in Portuguese and English: "BOVA onze" is found as BOVA11, "nove" matches 9, "nine" matches 9. Only for matching: the text is never rewritten from it. "um", "uma" and "one" are not folded.
• A close guess (fuzzy match) is no longer made when its words include an ordinary word the term does not have (D-083): "impressora que" is left alone instead of becoming "impressora 3D"; "BR Portness" → BR Partners and "fluxo de caixo" → fluxo de caixa still are. Exact, alias and phonetic matches are unaffected. On the fixtures no hit is lost and false positives go from 14 to 12 and 53 to 51 (frozen packs), 15 to 13 and 9 to 8 (bundled 0.3.7); about a third of the low-confidence marks go too.
• "to confirm N" is now "applied, please confirm N", in the summary and in the report (D-082): the close guesses are already in normalized.txt. The help and both guides say so, and that the review can undo them. The counts are the same.
• In the video's original language, a caption written by the channel is now taken before YouTube's automatic one, by fetch and as the menu's default (D-081, provisional, amending D-045). The order: the channel's own in the original language, the original audio's automatic one, other languages written by the channel, then automatic dubs and translations. --track still takes exactly the track named. None of the four fixtures has a manual track, so nothing changes for them.
• fetch without --lang takes the video's own language, as the site reports it, instead of always pt (D-080). When the site does not say: the video's only original-audio track, else its only manual track, else pt, with a line saying the language is unknown. --lang still wins.
• A link the downloader does not support says "This site isn't supported. Download the video another way and use 'A file on this computer'." instead of yt-dlp's "Unsupported URL" and a failure line (D-079).
• A site that gives no duration (Instagram) shows "duration: unknown" instead of "0min00s"; after transcribing, the audio's measured length is shown and written to meta.yaml, which now records the duration whenever it is known (D-079).
• fetch on a site with no captions at all says " doesn't provide captions for this video." instead of "there is no caption in 'pt' … (manual: none; automatic: none)"; the detailed line stays when there are captions in other languages (D-079).
• "How does this work?" in the sign-in list now speaks of any site: YouTube's check usually depends on the connection, Instagram's and Facebook's often on the video. Both guides say Kwai is not supported (yt-dlp has no extractor for it).
• Other sites (D-078). The sign-in is now "Sign-in (your browser's login)" and is used for every site, not only YouTube; a setting saved under the old name still works. A site that asks for a login (Instagram, Facebook, TikTok, X, Vimeo, private or followers-only videos) is said plainly: "Instagram asked you to sign in for this video.", or, when a sign-in is already set, "The browser you chose (Firefox) may not be signed in to Instagram. …", with the same ways out as YouTube's bot check (exit code 3).
• A link with no captions (most sites other than YouTube) goes straight to transcribing in the full build; the lite build says " doesn't provide captions for this video. The full build can transcribe its audio on this computer." and stops, with no error.
• The downloaded audio is now deleted once the transcription is written; Settings → "Keep the audio after transcribing", or fetch --keep-audio, keeps it in the run folder as <video id>.<ext>. Until now it stayed as audio.<ext>.
• fetch --list <url> is a quick check of a link on any site, without downloading: site, title, duration, caption tracks or none, whether there is audio, and whether the site asked for a login.
• After a review, only the questions that make sense (D-077): "learned layer written to …" appears only when the review changed it, and "Contribute what you taught the tool?" only when this review added or changed something there; answers that were all "only in this video" or "decide later" bring neither. "Add a term you noticed?" and "Open the folder?" are asked once each, in that order.
• The YouTube sign-in list shows the browsers found on this computer (D-077), including Firefox from snap or flatpak on Linux: Firefox, one "Chromium-based browser" row (a short list when there are several), Safari on macOS, a cookies.txt file, None, and "How does this work?". The "usually cannot be read" warning on Chromium-based browsers appears on Windows only.
• The review no longer rejects by Enter (D-072). A real review of 4tTmY8Buask rejected two terms when the user pressed Enter to skip a screen. The menu's review and normalize --review now stop once at each uncertain form, in the order of the video, with one question whose default is "Decide later": "Yes: this is (fix it from now on)", "Yes, but only in this video", "This is , said that way (recognise it, keep the text)", "No: this is not ", "Decide later". Only "No" rejects. Each stop shows the line before, the line with the form highlighted and the line after, the other times the form occurs, and a link that opens a YouTube video at that moment. Nothing is written until the end: a summary ("will remember … rejected … only in this video … left for later") and "Save these answers?"; no discards them all.
• "Only in this video" (D-073): kept in runs/<id>/local-decisions.yaml and applied when that run is normalized; never in the learned layer, a pack or a contribution. --confirm gains o for it.
• Dubbed videos (D-074): YouTube's automatic dubs add a caption track per dubbed audio. The video's own language is now the original: listed first and the default, the dubs named ", dubbed audio (automatic)", then English, then the rest. fetch --list shows them as dubbed:. With arrow keys, only the default row looks selected.
• No decision numbers in what the program says (D-075): help, menu, messages and command help no longer show "(D-0nn)".
• Linux full is a folder in a tar.gz, like Windows and macOS full (D-076): extract it and run transcript-normalizer inside the folder.
• Releases are gated on the tests: the release workflow runs the whole CI matrix (Linux, Windows, macOS; Python 3.11 and 3.12) on the tagged commit first, and builds and publishes nothing unless all six runs pass. Windows CI had been failing since 0.6.1 without stopping a release.

SHA-256

1d2ae7847e501cf0f608d62d407d2d5e06e20734e389fd722b6d28c91f710823  transcript-normalizer-0.6.5-linux-full.tar.gz
74c19df1f0820d67268997b65ab12f09d51ab4afbc4994b92293f4f0a79b968f  transcript-normalizer-0.6.5-linux-lite.tar.gz
89b5dc97b5732444b4e46a0ee9bcb96731f07f0ad7599249ece75da8f261fd18  transcript-normalizer-0.6.5-macos-full.zip
cec90bd64513e94b2cfe2b2864d6ed6ba03a2c1bdc9dceba4ff84c9e7654fceb  transcript-normalizer-0.6.5-macos-lite.zip
e1caaaf27f337f53af5548a412f8a472ec008de869c9213fbb8b99f2758abd74  transcript-normalizer-0.6.5-windows-full-portable.zip
484f5375276c27519dfb9d6a5d16551a346eaaee784dd0b4e78936054cc2e32b  transcript-normalizer-0.6.5-windows-full-setup.exe
539669150dbce549d660bcb6c9f1d1c6e63e98233a4eb2ef9f8310ad7acd9458  transcript-normalizer-0.6.5-windows-lite.exe
3ac863b6f14a612ab268de9785328db99819ea850798af9126ad17ee5b5111f4  transcript_normalizer-0.6.5-py3-none-any.whl
20d0bcc58929d951828568f76719b669227800c1f034225d5df7abe8747dbb46  transcript_normalizer-0.6.5.tar.gz