CaptionLab — Changelog
======================


v1.0.6 — September 2026
─────────────────────────────────────────

  ✂️   Remove punctuation
    • New switch in the settings — strips full stops and commas from the
      finished captions, the usual look for short-form video
    • Timing and line breaks are unaffected, so nothing moves on the timeline
    • Works in every language, including Arabic, Hindi and Chinese, which use
      punctuation marks of their own

  ⚡   GPU and memory
    • The graphics libraries are now verified by loading them, so a
      half-installed CUDA can no longer report success and fall back to the
      processor without saying so — the usual cause of "it works, but slowly"
      on Windows
    • The model releases graphics memory between jobs instead of holding
      several gigabytes, leaving it free for playback and previews
    • Only one engine runs at a time. A second copy could stay loaded after
      reopening the panel or switching Windows users, and switching models
      stacked them in memory

  🎓   Education licences
    • A separate version for universities, schools and studios: one key for an
      agreed number of computers, and nobody using them has to enter anything
      — no accounts, no activation screen
    • For pricing and a quote, write to editlabtools@gmail.com

  ⚙️   Settings menu
    • Everything that used to sit in different corners is now under the gear:
      repair the engine, open the install log, what's new, request a feature,
      leave a review, help
    • Repair engine re-checks every part of the AI engine and downloads only
      what is missing or damaged — your models are not re-downloaded. Until
      now a broken install had no route back except deleting everything
    • "Check for updates" at the bottom of the panel, on click only
    • The CaptionLab logo takes you back to the main screen from anywhere

  🐛   Fixes
    • A failed GPU download no longer aborts the whole install — setup
      finishes and falls back to the processor. If a download is blocked, it
      tries other routes automatically
    • If an antivirus quarantines CaptionLab's files, the installer now names
      it as the cause and lists what to whitelist, instead of only reporting
      that files went missing
    • Interrupted downloads can no longer be mistaken for finished ones on the
      next attempt
    • "Max words" is now a strict limit. A caption could come out one word
      over the number you set — including when a number like "39 dollars"
      arrived as a single word from the model — and no longer can


v1.0.5 — August 2026
─────────────────────────────────────────

  🎬   Free 3-day trial
    • The complete plugin before you pay anything — every model, every
      language, real subtitles in your timeline. No card, no account
    • You get your key BEFORE installing anything, so you can set up at your
      own pace. The three days start at your FIRST TRANSCRIPTION, not at
      download — installing the engine and pulling a multi-gigabyte model can
      take an evening, and none of that comes out of your trial
    • Buying afterwards replaces the trial in place — everything you already
      downloaded stays exactly where it is

  🇲🇲   A model built for Burmese
    • Burmese is one of the languages the standard models handle badly: on
      real footage they produced the wrong script entirely, or repeated the
      same syllable for minutes
    • A Burmese-specific model now sits in the Models tab and produces real,
      readable Burmese on the same clips
    • It exists because a customer asked for a refund when his language
      didn't work. If yours doesn't either, say so — editlabtools@gmail.com

  🔧   Installs that repair themselves
    • A half-finished download used to be trusted forever: every retry saw
      the file was "there" and skipped it, so the same broken install came
      back however many times you clicked
    • Each piece is now checked to see it actually RUNS, and re-fetched if
      not. Stalled downloads time out and retry instead of hanging

  🌍   Interface translated into 36 languages
    • Every new screen — the trial, the models, the errors — reads in your
      own language rather than falling back to English


v1.0.4 — August 2026
─────────────────────────────────────────

  ⏱   Caption Timing — rebuilt
    • Caption starts are now measured against the audio itself instead of
      trusting the model's own timestamps. Captions no longer appear while
      the speaker is still silent
    • On a real news package this cut total misalignment by ~60%, and made
      Large-v3 as accurate on timing as Large-v2
    • Works after long pauses, B-roll with room tone, and music beds — the
      cases where captions used to show up seconds early
    • Captions now hold slightly longer after the last word instead of
      vanishing the instant the speaker stops

  ✂️   "Fixed words" is now "Max words"
    • The setting is a ceiling, not a quota. Set 4 and a caption will be 3,
      2 or even 1 word when that is where the line logically ends
    • A caption is never split across a sentence boundary, and breaks at
      commas and clause boundaries
    • Short binding words ("not", "the", "in" and their equivalents in 20
      languages) are never left stranded alone on screen
    • Renamed across all 36 interface languages

  📝  Default captions — proper grouping
    • The default mode previously wrote the model's raw segments straight
      to the SRT. Those end where decoding stopped, not where a sentence
      ends, which meant a lot of manual re-cutting
    • Captions now respect sentence and clause boundaries, line length and
      reading speed
    • Hyphenated words ("self-aware") are no longer split across two
      captions

  🔊  Quiet speech no longer skipped
    • Audio recorded well below the level of the rest of a video — field
      interviews cut against a loud studio voice, for example — was being
      dropped silently. Levels are now normalised before transcription
    • Measured: speech at -36 dB and below was previously lost entirely

  🐛  Fixes
    • Fixed the first word of a new sentence being attached to the end of
      the previous caption
    • Fixed hallucinated repeated phrases on quiet or noisy passages
    • Mac archive now ships the signed .zxp, as intended


v1.0.3 — July 2026
─────────────────────────────────────────

  🧠  Model Manager
    • New "Models" tab — choose which Whisper model is active, download or
      delete models right from the panel
    • Added Large-v3 (better on noisy audio & accents) and Large-v3-turbo
      (~6x faster) alongside the existing Large-v2 default, plus a
      lightweight Small model for weaker machines or quick drafts
    • Shows real download progress and total disk space used by models
    • Already-installed models (from a previous version) are recognized
      automatically — no unnecessary re-downloading

  ✂️   Caption Length
    • New "Fixed words" caption mode — choose 1-8 words per caption instead
      of the default sentence/pause-based grouping
    • Ideal for vertical video (Shorts/Reels/TikTok) captions that pop on
      screen in sync with speech
    • Works on Mac (Apple Silicon and Intel) and Windows

  🌐  Localization
    • The new Models tab, caption-length controls, and all related buttons
      are fully translated across all 35+ supported interface languages


v1.0.2 — July 2026
─────────────────────────────────────────

  🎬  Multicam & Merged Clips
    • Fixed "No media found" / "No clip selected" on multi-camera sequences
      and clips with externally synced audio
    • Audio is now rendered directly from the timeline instead of chasing
      each clip's source file, so it works for multicam sequences, Merged/
      Synchronized clips, nested sequences, and clips with audio effects
      or gain applied
    • Status message now reflects what's being transcribed (clip, multiple
      clips, full sequence, or multicam) instead of the model name


v1.0.1 — July 2026
─────────────────────────────────────────

  🪟  Windows
    • Fixed a false "Setup failed" error during AI engine installation caused
      by a harmless pip warning being mistakenly treated as fatal
    • Background engine now auto-starts on login (Windows Task Scheduler)
    • Manual "Start server" fallback for cases where auto-start doesn't fire

  🔧  Reliability
    • Build fingerprint (BUILD_ID) added so the panel can detect a stale
      server process after a reinstall/update
    • Machine ID now persists in a shared file instead of per-app localStorage,
      fixing device-limit false positives across multiple Premiere versions
      on the same computer


v1.0.0 — June 2026  (Initial release)
─────────────────────────────────────────

  ✨  Features
    • Auto-generate subtitles from audio in Premiere Pro
    • 99 languages supported (vs. 16 in Adobe's built-in)
    • Transcribe in source language OR translate to English
    • Works on Selected Clip or Full Sequence
    • Respects gaps between clips on the timeline
    • Word-level timestamps for clean SRT segmentation
    • Auto-detect spoken language

  🔒  Privacy
    • 100% offline — audio never leaves your Mac
    • No API keys, no accounts, no telemetry
    • Whisper large-v2 AI model runs locally

  🌐  Localization
    • Interface available in 35+ languages
    • Auto-detects macOS system language on first launch

  ⚙️   First-run setup
    • One-click installation of Python runtime, ffmpeg, and AI model
    • No terminal, no admin password, no Homebrew required
    • Portable Python — fully sandboxed, easy uninstall
    • Real-time progress with download size shown in MB

  🎨  UI
    • Native dark theme matching Premiere Pro
    • Pulse indicator on Ready state
    • Cmd+Enter keyboard shortcut to generate
    • Adaptive layout for narrow panel widths

  💻  System
    • Apple Silicon: mlx-whisper backend (optimized for M-series)
    • Intel Mac: faster-whisper backend
    • Premiere Pro 2022 or later supported


────────────────────────────────────────────────────────────────────

Made by Edit Lab Tools  ·  editlabtools@gmail.com  ·  captionlab.pro
