Small updates.
Better videos.

Adjust word and phrase timing with timeline handles, drag to move, arrow keys, or exact start and end fields. Undo and Reset are available.

Preview video at 1×, 1.1×, 1.25×, or 1.5× speed while preserving speech pitch. Export speed stays unchanged.

Camera audio is captured on the audio thread and buffered for AAC encoding, with stereo support and a final audio flush when recording stops.

The transcription progress bar and 100% Local badge now use the yellow brand colors.

Long videos now prepare one audio window at a time, so transcription can start before all windows are prepared. Completed windows no longer retain audio features, and the final transcript reuses the last merged result.

Timing edits preserve original video cut positions and phrase grouping. Preview and subtitle exports use edited times, and timing Undo preserves text and style edits.

Subtitle timing edits survive transcription updates. SRT and WebVTT timestamps round correctly to milliseconds, including minute boundaries.

Stopping audio capture and draining the encoder now have a shared 15-second limit. Failed or cancelled audio finalization releases capture resources and reports an error instead of waiting indefinitely.

Drag subtitles to move them and use corner handles or keyboard controls to resize. Keep placement local to one caption or apply it globally.

Stacked portrait layouts for two-person conversations, with face tracking, top/bottom swap, and adjustable face zoom.

A zoomable video timeline with thumbnails, subtitle segments, precise scrubbing, playback controls, and quick access to framing settings.

Silence removal with conservative, default, and aggressive detection, plus automatic face-centered zoom cuts.

Camera recording with camera and microphone selection, a recording timer, and a review step before editing.

Redesigned the landing page and editor with warm backgrounds, yellow accents, clearer controls, and dedicated Style, Subtitles, and Video panels.

Rebuilt mobile editing around a compact tool dock and a preview that adapts to the available screen space and keyboard.

Best is now the default export quality, using quality-prioritized H.264 encoding and 256 kbps audio. Mobile MP4 exports use up to 6 Mbps, 24 fps, and a 1280-pixel long edge.

Word editing is available in both word and phrase modes, with a movable desktop panel, mobile controls, and an emoji picker that stays visible.

Added clearer model download and transcription progress, automatic WebGPU-to-CPU fallback, and more useful retry messages.

Refreshed the changelog to match the new brand and synchronized the app and package versions.

Caption placement, text edits, emojis, hidden subtitles, and removed sections now survive incoming transcription updates.

Word font, size, color, knockout, emoji replacement, and emoji overlays are honored during export, including when phrase emphasis is off.

Stacked exports retain portrait framing for square and non-9:16 portrait sources, with encoder-compatible dimensions.

Reduced export memory duplication and explicitly release canvas buffers and decoder resources on completion, cancellation, and leaving the editor.

Kept the person visible in background-removal previews and improved background-processing cancellation and media cleanup.

Improved subtitle case handling, active-word background colors, preview font scaling, and watermark sizing.

Improved handling of missing or silent audio, media loading and seeking, skipped sections, and cancelled processing.

September 8 maintenance: fixed low-sample-rate AAC export compatibility on Windows, retained download URLs long enough for the browser to read them, and closed audio samples during cancellation.

September 8 maintenance: stabilized recording timers and component state, refreshed install-banner behavior, updated dependencies and lint configuration, and added media regression tests and Next.js contributor guidance.

Double audio extraction / transcription triggered by clicking Transcribe twice — added ref guard in useTranscription to prevent concurrent calls

1-2 second idle gap between language modal closing and processing overlay appearing — status now set synchronously in the same render batch as modal close

Aspect ratio tabs now use rectangle icons (horizontal/vertical) instead of plain text, with labels on desktop

Mobile Styling/Edit tabs restyled with primary color active state and bolder font to establish clear visual hierarchy above the Word/Phrase sub-tabs

Bottom fade shadow on mobile scroll area to hint that more content is available below

Removed drag-to-pan and fit/crop toggle for portrait-in-landscape — face tracking handles intelligent cropping; cleaned up all related code from video-upload, main-app, and export hook

Font subset error from fonttrio install — removed invalid 'menu' subset from Jost, Rubik, and Roboto_Mono declarations in layout.tsx

Disabled React Strict Mode to prevent dev-only double-mount side effects (double worker init, double audio extraction)

Full mobile-first layout: viewport-locked page with no page scroll, inline Styling/Edit tabs below the download button with internal scrolling, and an expand button for full-screen editing

Converted entire codebase from hardcoded Tailwind colors to semantic theme variables (bg-primary, text-foreground, bg-muted, etc.) so shadcn presets apply correctly

Moved custom utility functions out of lib/utils.ts into lib/transcript-utils.ts so shadcn preset installs no longer overwrite app code

All mobile controls now use shadcn Button component with proper variants and sizes for consistent touch targets and interaction states

Regenerate Subs button (formerly Change Language) now opens the language modal without auto-starting processing, letting users pick a different language first

Confirmation dialog (shadcn AlertDialog) when uploading a new video to prevent accidental loss of current subtitles and styling

Confirmation dialog for background removal on videos longer than 1 minute warning about processing time

Cancel button with live progress shown inline during background removal processing

Slim progress bar below controls during background removal on mobile

Mobile bottom sheets (Styling/Edit) no longer blink and fail to open — replaced dual-SheetContent pattern with separate Sheet components

Language selection modal now properly resets state when reopened, preventing stale Processing state from previous transcription

New Formal subtitle preset using Playfair Display with white text, a subtle dark outline/shadow treatment, and stronger preset button contrast for light styles

Separate active-word recolor emphasis control with a configurable highlight color, supported in both preview and export rendering

Default subtitle styling now starts with the Formal look, 3 max words per line, display-on-spoken disabled, active-word recolor enabled, and a -10px vertical offset

Active-word emphasis is now split into independent resize and recolor controls so each effect can be toggled separately

Export quality selector — choose between Medium (smaller file, closer to source bitrate) and High Quality before downloading

Export pipeline now uses OffscreenCanvas, alpha-free contexts, and StreamTarget for reduced memory usage and faster frame rendering

Default export quality lowered from High to Medium so output file sizes match the original video more closely

Face position analysis before 9:16 export now runs at 2fps instead of 5fps (2.5x faster) and shows a live progress percentage

Mobile MP4 export now uses more conservative compatibility settings during the playback-freeze investigation, including explicit AVC settings on mobile, lighter mobile export caps, and on-screen export diagnostics for codec, MIME type, bitrate, and source-track details

HEVC-source export startup now streams audio samples alongside video rendering and resolves the output MIME type asynchronously so the ongoing 0% stall investigation is no longer front-loaded on whole-file audio preparation

Canvas context state (transforms, composite ops) no longer resets every frame during compositing export — blur, foreground, and mask canvases are sized once and cleared instead of recreated

Subtitle chunk lookup during export uses binary search instead of linear scan, eliminating per-frame overhead on long videos with many subtitles

Subtitle preview text on mobile now scales relative to the video container width instead of the viewport, with a readability boost on small screens so text stays legible even on compact previews

Export subtitle positioning now matches the preview baseline, so vertical offset and default bottom placement render at the expected height in downloaded videos

Export frame extraction now clears the canvas every frame, falls back to the source video when a decoded sample is unavailable, and detects HEVC phone-camera sources so they can use sequential decoded-frame reads instead of sparse timestamp lookups during export

Installable Progressive Web App support — Based Subtitles can now be added to the home screen and launched in standalone app mode

Generated branded app icons using the existing BS monogram for Android, desktop install prompts, and Apple touch devices

Install banner — shows a native install prompt on supported browsers and Add to Home Screen instructions on iPhone/iPad Safari

App Router metadata now includes a web app manifest, Apple web app settings, and service worker wiring required for installation

Split subtitle controls are now always visible in the styling panel, so top/bottom and left/right layouts can be selected without depending on 3D depth mode visibility

Display-on-spoken subtitles in dynamic 3D preview and export now keep a stable layout — future words stay reserved in place and no longer cause earlier words to shift when they appear

Per-word editing chips now remain available for single-word phrase subtitles, so brief isolated words can still be selected and styled from the bottom editor bar

Split subtitle mode now preserves per-word emoji replacement and emoji overlay styling in both preview and exported video instead of dropping those overrides on split lines

Vertical offset slider (±50 px) to fine-tune subtitle position above or below the default — applies to preview, canvas compositing, and exported video

Vertical offset now correctly applied during video export — previously the setting was ignored and subtitles always rendered at the default position

"Hide inactive" toggle in the transcript sidebar — hides both skipped segments (no audio) and hidden-subtitle segments (audio plays, no text) to keep the sidebar clean

Project is now open source — source code available on GitHub at github.com/deifos/basedsubtitles

Streaming transcription — partial results appear chunk by chunk as Whisper processes the audio instead of waiting for the full video to finish

Live subtitle preview updates in the transcript sidebar while transcription is still running

Non-blocking UI after the first chunk arrives — blocking overlay replaced by a slim in-header progress indicator so you can see partial subtitles in real time

Transcribing banner in the transcript sidebar (amber, with spinner and progress bar) pinned below the scroll area so it stays visible while reading results

Transcribing banner now shows the current timestamp being processed (e.g. Transcribing 0:42 · 73%) so you can track progress without scrolling

Long video support is now reliable — transcription uses the same 30s / 5s stride chunking that the Whisper reference implementation uses, with stride-aware multi-chunk merging via _decode_asr

Whisper worker now replicates the pipeline's internal _call_whisper loop (model.generate() per chunk + tokenizer._decode_asr() for merging) instead of calling the high-level pipeline function, giving identical accuracy with streaming output

Transcription was silently skipping sections of audio in videos longer than ~30 seconds — caused by sending the full audio in one call without stride-aware chunking

Multi-chunk merging was incorrectly concatenating chunks instead of filtering stride overlap regions — fixed by enabling return_timestamps so timestamp tokens are present for _decode_asr's stride filtering logic

Subtitle file downloads used a generic filename (subtitles.srt) regardless of the source video — now uses the video filename as the base (e.g. my-clip.srt)

Removed 'Add Subtitle' manual entry form from the transcript sidebar — no longer needed now that transcription accuracy covers the full video without gaps

Split subtitle mode — phrase text is divided in two and placed around the person's head. Two modes: Above/Below (works on any ratio) and Left/Right (widescreen only)

Left/Right split positions text at eye level (~38% from top) with a gap on each side of the face so text has room to move as the person shifts

Per-phrase frozen face position for split subtitles — text is anchored to where the face was at the start of each phrase and stays fixed for its duration, avoiding distracting mid-phrase drift

Face tracking auto-starts when a split subtitle mode is selected and stops when set back to Off — no manual toggle required

Display on Spoken — all phrase words appear dim at phrase start and each word lights up to full brightness when spoken. Layout is stable from the first word (no shifting as words appear)

Display on Spoken works in all render paths: DOM preview, canvas compositing, and video export

Display on Spoken combined with Split Subtitle mode — each half renders words individually with dim/bright per word

Split Subtitle section is always visible in the styling panel — no longer gated behind face tracking being active

Applying a style preset no longer resets the font size — presets change color, border, and font family while preserving the user's chosen size

Default style updated to green (#00FF41), medium font size, no background, and Display on Spoken enabled

Face tracking was not applied during 9:16 video export — the buildExportTimeline function was missing from the useCallback dependency array, causing the export to always use a stale closure where it was undefined

Exported face tracking had a slight delay compared to the preview — added a 150ms look-ahead offset when sampling the face timeline during export to compensate for EMA smoothing phase lag

Export failed with odd-dimension error on 4K sources — 9:16 crop of 2160p height produces 1215px width (odd), which H.264 rejects. Dimensions are now rounded down to even numbers

Font stroke/border was filling inside the text at higher widths — canvas strokeText draws centered on the glyph outline so half went inward. Now draws stroke at 2x width before fill, and CSS preview uses paint-order: stroke fill

Border width control replaced with a 0–20px slider instead of a dropdown with only 4 options (None/Thin/Medium/Thick)

Default subtitle style now uses 10px white border stroke for better contrast and readability on any background

Removed text fade-in effect (letter-by-letter reveal) — timing was unreliable, words would disappear before fully revealing on short chunks

Exported video was low quality on mobile — the export pipeline creates multiple full-resolution offscreen canvases that exceeded mobile browser memory limits, causing silent quality degradation. Export resolution is now capped at 1080p on mobile to stay within canvas memory budgets

Black video export on mobile browsers without WebCodecs decoding support — when canDecode() returned false, no video frames were drawn. Added a fallback that seeks the video element and draws each frame directly

Color pickers caused heavy lag while dragging — all 5 color inputs (per-word color, text color, background color, border color, solid background) fired onChange on every pixel movement, triggering full state updates and canvas re-renders 10–50 times per second. Now debounced with immediate visual feedback via local state

Camera recording showed black screen on mobile — video element didn't exist in the DOM when the stream was attached; now re-attaches via useEffect when the preview mounts

Video not playing after scrubbing on mobile — pausing the video on drag start then calling play() on release failed because pointerup on range inputs isn't a trusted gesture on mobile; removed pause/resume entirely

Video freezing when scrubbing on mobile — was setting videoEl.currentTime on every drag frame, overwhelming the decoder; now only seeks once on pointer release (mediabunny pattern)

Subtitles showing stale position during seek drag — now calls onTimeUpdate during drag so subtitles and transcript stay in sync with the scrub position

Video not buffering fully on mobile — added preload="auto" so mobile browsers load the complete video instead of lazy-loading

Replaced <input type="range"> seek bar with custom div-based progress bar using pointer events and setPointerCapture — eliminates all mobile range input quirks and provides reliable touch tracking even when finger drifts off the bar

Progress bar fill and time display update via refs during drag (zero re-renders while scrubbing, Vercel React best practice: useRef for transient values)

Added keyboard support (left/right arrows ±5s) and ARIA slider attributes to custom progress bar for accessibility

Video timeupdate events are ignored during seek drag to prevent the video's stale position from overriding the user's scrub position

Per-word emoji size slider — scale each word's emoji independently from 50% to 200% (appears in word style popover when an emoji is set)

Compact word style popover on mobile — collapsible icon tabs (Font, Size, Color, FX, Emoji) overlay the video without blocking subtitles or requiring scroll

Word editing standardized to chip bar below video — works in all modes (3D on/off), replacing direct click-on-word in the preview

Word chips restyled with dark background for better visibility

Min words per line lowered from 3 to 1 — allows single-word-at-a-time subtitle display

Emoji size is now per-word instead of global — moved from styling panel to the word style popover

Emoji overlay size was not applied during video export — now correctly uses per-word emojiScale in all render paths

Per-word emoji replace — click any word and pick an emoji to replace the word text entirely with that emoji

Per-word emoji overlay — place an emoji above any word while keeping the text visible underneath

Emoji picker integrated into the per-word style popover with search support

Emoji replace and overlay render in DOM preview, canvas compositing (3D depth mode), and video export

Emoji size scales with the word's font size — enlarging a word also enlarges its emoji

Word style popover in 3D mode did not update when clicking a different word badge — popover now resets and shows the correct word's settings

Branding watermark — toggleable "basedsubs.getbasedapps.com" text in the bottom-left corner, rendered in preview and baked into exported videos

Letter-by-letter text fade in — characters reveal left-to-right within each word with staggered timing, similar to VEED-style animated captions

"Text fade in" toggle in the styling panel — when off, text appears and disappears instantly with no fade effects

Letter-by-letter reveal works in DOM preview, canvas compositing (3D depth), and video export

Word and chunk fade effects are now gated behind the text fade in toggle instead of always on

Active word emphasis is now off by default

Mobile camera starting with black screen — autoPlay not reliable on mobile, now explicitly calls play() after attaching the stream

Flipping camera showed "camera in use by another app" — the useEffect re-fired openCamera on facingMode change, causing two getUserMedia calls to race for the same device

Camera defaulting to back camera on some Android devices — now uses exact facingMode constraint instead of a preference hint the browser can ignore

Mobile styling and edit panels could not be switched directly — had to close one before opening the other. Bottom bar buttons are now toggles with mutual exclusion

Mobile drawer panels cut off at the bottom — content now fills the full available height down to the navigation bar

Mobile styling panel couldn't scroll to the last items — scroll area now accounts for the fixed bottom navigation bar

Letter-by-letter text fade in now works in 3D depth mode — both behind and front text layers support per-character reveal and chunk fade-out

Mobile transcript panel height was capped at 384px (max-h-96) — removed the cap on mobile so the list fills the drawer

Removed redundant "Subtitles behind person" toggle — its functionality is fully covered by the "Dynamic depth 3D" toggle

Knockout text effect — per-word style option that makes the text shape reveal an inverted/negative version of the video behind it

Knockout toggle in the word style popover under a new 'Effect' section

Knockout renders in DOM preview (mix-blend-mode: difference), canvas compositing (3D depth mode), and video export

Lilita One font — thick, bold display font with a magazine-cover aesthetic, great for knockout effects

Footer 'Powered by' row no longer overflows on small screens — items wrap and bullet separators hide on mobile

Video freezing on mobile after transcription — unhandled play() promise rejections from autoplay policy now caught on all play/pause controls

Toast notifications for background removal errors — shows a clear message when WebGPU/WASM is unavailable instead of silently failing

Camera recording — record video directly from your camera (front or back) with a 'Record video' button on the landing page

Records straight to MP4 using MediaBunny (H.264 + AAC) — no WebM conversion needed

Camera flip button to switch between front and back cameras on mobile

Review screen after recording with re-record and 'Use this video' options

Custom player controls (play/pause, seek bar, time, mute) shown consistently in all modes

Click anywhere on the video to play/pause

Umami analytics integration

Powered by Transformers.js and MediaBunny attribution with logos in the footer

Removed native browser video controls — custom controls prevent subtitles from covering playback buttons

Landing hero tagline updated from 'Free forever' to 'Free'

Selecting the 500MB model caused a blank screen — base model preload raced with the user's model choice, resolving the wrong download promise

Added load ID tracking so stale model 'ready' messages from previous downloads can't resolve the wrong promise

Cancelling the language modal on a fresh upload now returns to the landing page instead of showing a dead-end video screen

Subtitle text appearing/disappearing no longer causes layout jumps — placeholder space is always reserved

Word chip bar in compositing mode no longer causes layout shifts between phrases

Video export with 3D depth enabled was re-running MODNet AI inference on every frame at 30fps (~1,800 calls per minute of video) — now uses pre-computed 5fps mask cache, reducing export time from ~15 minutes to under a minute

Export with 3D depth consumed ~2GB+ RAM from per-frame 8MB getImageData allocations and data copies to the AI worker — eliminated entirely by using cached masks

Word chip bar — clickable word buttons appear below the video when 3D depth mode is active, letting you select any word for per-word styling

Words with custom style overrides are highlighted with an amber tint in the chip bar

3D depth mode now auto-enables when background removal starts — no extra toggle needed

Per-word selection in 3D/bg-removal mode uses the reliable chip bar instead of canvas click coordinates

Clicking anywhere on the canvas in 3D mode no longer incorrectly triggers word editing

Hidden and disabled words can no longer be selected for editing

Major memory leak in video export — frame pixel data (~8MB) was allocated per-frame instead of reusing an offscreen canvas via GPU blit

AudioContext not closed on error during export, leaking audio memory

Reusable canvases and buffers from export were never released after completion, staying in memory indefinitely

Font mapping table (24 entries) extracted to a single module-level constant — was duplicated in 5 functions and recreated per-frame during export

Removed dead hitTestOnly prop from VideoCaption and unused renderDynamicWordToCanvas function

Per-word custom styling — click any word in the video preview to override its font, size, and color independently

Word style popover with font family picker, size multiplier (50%–200%), color picker, and reset button

Selected word highlight with yellow outline in the video preview

Per-word overrides render correctly in both the live preview and exported video

Preset buttons now render with their actual font — Bangers for Green, Permanent Marker for Gold, Outfit for Subtitle, Bebas Neue for Gamer

Preset buttons show the correct font weight so you can preview the style before applying

New landing page with hero section, feature highlights, how-it-works flow, local processing showcase, and call-to-action

Full-page drag-and-drop — drop a video anywhere on the landing page to get started

Drag-over visual feedback on the dropzone with amber highlight

Changelog page at /changelog with version history

"Buy me a coffee" floating button

Site footer matching getbasedapps design with version link

Redesigned feature cards with playful visual representations — waveform bars, layered text, font samples, language globe

Subtitle style presets now use dynamic fonts (Bangers, Permanent Marker, Bebas Neue, Outfit) instead of generic system fonts

Default subtitle font size changed to Small for a cleaner look

Subtitle styling panel now uses Outfit font with consistent rounded-lg buttons and amber accent colors

Style presets now visible in dynamic (3D) mode for consistent styling

Subtitle preview pinned above the scroll area so it's always visible while adjusting settings

Editor buttons and panels updated to match the new landing page design

Major memory leak — Whisper model (~1GB) now freed after transcription completes instead of staying in memory

Major memory leak — background removal model (~500MB-1GB) now freed after processing completes

Mask data was duplicated in both a ref and React state (~2x memory usage) — removed unused state copy

Canvas render loop now skips redundant draws when video is paused, saving significant CPU

ImageData allocation (~8MB) was created every frame at 60fps — now reused across frames

Temporary canvases in video export now reused across frames instead of recreated per-frame

Removed purple color accents — replaced with amber tones throughout

AI-powered subtitle generation using Whisper.js — runs 100% locally in the browser via WebGPU/WASM

Multi-language transcription supporting 100+ languages with selectable Whisper model sizes (tiny, base, small)

Background removal with AI person segmentation — processes video frames locally without any uploads

Dynamic 3D subtitles — place text behind or in front of people with depth effects

Follow-word mode for dynamic subtitles — text tracks the position of the spoken word

25+ Google Fonts including Bangers, Montserrat, Bebas Neue, Poppins, Oswald, Anton, Fredoka, Permanent Marker, Pacifico, and more

Full subtitle styling controls — font, size, weight, color, background, border, drop shadow, word emphasis

Word-by-word and phrase display modes with configurable max words per line

Subtitle position control (top, middle, bottom) with adjustable Y positioning for dynamic mode

Video export with baked-in subtitles using Mediabunny — downloads MP4 with subtitles permanently rendered

Landscape (16:9) and portrait (9:16) aspect ratio modes with zoom toggle for portrait

Editable transcript sidebar — click any segment to modify the transcribed text

Language selection modal with model size picker before transcription starts

Mobile-responsive design with drawer navigation for styling and transcript editing

Works offline after first load — all AI models are cached in the browser

Free to use — no sign-up, no watermarks, no data collection