scenaristo
Download free
Changelog

Everything that shipped, release by release.

Scenaristo ships often and in the open. This page collects every release since the first signed build, written out in full: the features, the reasons behind them, and the fixes that made them behave. Newest first.

Download v0.4.2Installers on GitHub
Jump to a release

v0.4.2

Latest
SpeakersTakesTranscription

The speakers release: identification that can tell voices apart, a way to say how many there are, and words that stop waiting for their labels.

Speaker identification that earns the checkbox

Identify speakers runs on a new engine. On a real three-speaker call, the old one pinned 93% of the words on a single voice, and no amount of tuning moved that by more than a point; the new one tells the speakers apart, and its model download is less than half the size (24 MB against 50.6). You can also rename a speaker right in the transcript now, where you are already reading, instead of reaching for the rail card.

Say how many voices there are

A "Number of speakers" field sits beside Identify speakers when you transcribe, and the Speakers card gains the same control after the fact, with its own run button, live progress, and a Stop. Trying another count re-runs only the speaker pass, not the whole transcription, and the names you have typed are carried across wherever the match is unambiguous. When the engine finds fewer voices than you asked for, it says so plainly, and gives the honest reason: this audio may only hold that many voices that can be told apart.

Words first, speakers behind them

With Identify speakers on, the transcript used to wait for both passes, even though the words were finished long before the labels. Now the words appear the moment transcription completes and the speaker labels land behind them. On a ten minute test recording, that is the difference between reading at 15 seconds and reading at 98. The transcript is fully editable the whole time: scroll position, selection, and an edit in progress all survive the labels arriving, and if the speaker pass fails you keep a good transcript with a quiet retry instead of losing the run.

Take review moves into the transcript

Take groups now review where you read. The takes proposed for the cut are painted with a dashed strike, the keeper gets a quiet keep wash, and each group anchors a card in the margin: a preview first, expanding in place into the full review, where you audition each take, pick the keeper, and accept or dismiss. On a narrow pane the cards collapse to a small count at the edge and float over the text when focused. The rail’s Takes card stays as the overview: the scan lives there, and each group is a row that jumps you to it. An accepted group drops its paint but keeps its card, for the change-your-mind path.

Also in this release

  • The filler cleanup and silence removal toolbar buttons get real icons.
  • Opening a project is lighter on the wire: every setting arrives in one read instead of twelve, the startup read burst waits until the project is open, and one shared clip cache replaces one per caller.
  • The "audio sounds like X" model-mismatch banner is gone.

v0.4.1

Takes

The app learns to spot a retake: say the line four times, keep the best one, and let it strike the rest.

Keep the best take

Record the same line until it lands, and the app can now find every attempt. A scan proposes groups of repeated takes; each group gets a review card where you audition the takes side by side, pick the keeper, and accept, which strikes the others in one go. Dismiss a group and nothing is touched: the proposals are suggestions, not edits, until you say so.

v0.4.0

TimelineCrossfades

Every cut can carry a crossfade, and the timeline gets easier to read and to drive.

A crossfade on any cut

Every seam between kept material can now carry an audio crossfade. Add or remove one on the seam itself in the timeline, and playback and export render the identical fade, so what you hear while editing is what the file ships. Whether a fresh cut starts with a crossfade is a project setting, decided once rather than remembered per cut, and the fades you place are stored with the project and restored when you open it again.

A timeline you can drive

Drag anywhere on the lane stack to scrub the playhead, instead of aiming for the ruler. Each track’s lane takes a height sized to what it shows, and the ruler and lane content fit the main pane instead of running past its edge.

Also in this release

  • macOS 14 Sonoma is now the minimum system version.
  • The timeline’s cut layer is restored when a project is reopened.
  • The project title holds to one line, and its duplicate in the toolbar is gone.
  • Opening a media container moved off the dispatcher’s lock, so it no longer stalls every operation queued behind it.
  • nanoid is bumped past a published advisory.

v0.3.3

TimelineFillers

The timeline stops showing the recording you imported and starts showing the video you are making.

The timeline shows the edit

The video and audio tracks now lay out on the edited timeline instead of the source. Cut material is simply not drawn: each kept segment becomes its own clip block, so every cut reads as a seam between clips, and the waveform closes up at each seam instead of dimming across it, refining clip by clip as you zoom in. The ruler, the duration chips, and the playhead all read output time, and struck words vanish from the text track. Because the layout comes from the same projection playback uses, the timeline, the reader, and what you hear agree by construction.

Fillers, in Spanish

The filler pass picks its vocabulary from the language of the recording, detected automatically or taken from the language you chose when transcribing. A Spanish recording gets a Spanish vocabulary, so its fillers are flagged as readily as an "um".

Also in this release

  • The time ruler draws at most one tick per text line and skips a boundary that a long silence swallows.
  • A project location you chose is kept, and media paths are rebased when a project is copied.

v0.3.2

SpanishTimecodes

The interface learns Spanish, and the transcript’s timecodes grow up into a real ruler.

The app, in Spanish

The whole interface is available in Spanish, chosen in Preferences, and not just the webview’s half: the native menu bar is localized from the same catalogs, so File and Edit follow the language too. Numbers format per locale, and the marker import surface arrived translated.

A ruler for the edited timeline

The output timecodes, rethought: instead of markers sitting inline with the words, a vertical ruler runs in the left gutter, with a tick at each thirty second boundary of the edited timeline placed at the exact text line where that boundary falls. Minute ticks carry a mono MM:SS label, the first kept word reads 00:00, and the ticks reshuffle live as you strike and restore material. The question it answers is unchanged: how long is the thing you are actually making.

Also in this release

  • Menu items that carry the app’s name keep "Scenaristo" as is in every language.

v0.3.0

LabelsReactionsMultitrackCorrections

The long-recording release: mark up a stream in one pass, let the app find the moments you shouted at, and tell it which of your audio tracks is actually you.

Labels, for the pass where you decide what matters

Select any stretch of transcript and label it, with ⌘⇧L (Ctrl+Shift+L on Windows) or the selection popover. Give it a name, add a note if you want. Labels sharing a name are a group, so the Labels card lists them together, with a filter over names and notes, click to jump, inline rename, and delete. Unlike the single clip, labels pile up: they may overlap and nest, and tagging one sentence twice is the normal case. The selection survives an add, so the next tag is one keystroke. Labeled words carry a faint tint in the reader, and the ranges follow your edits: cutting inside a label shrinks it, and deleting all of its text leaves it flagged as orphaned rather than quietly disappearing.

The app finds the moments you reacted to

A new Reactions card scans your voice track for the moments your own energy spikes: the screams, the laughs, the hype. These are the highlights transcript search can never find, because there are no words in them. Each proposal comes back as a clip of the length you pick (15, 30, or 60 seconds) with the moment two thirds of the way in, so the payoff lands in the final third. Nothing is placed for you: review a proposal, then dismiss it or accept it, and accepting labels the whole clip “reaction” in a single undo step so your label triage takes over from there. The scan runs when you ask for it, with a real percentage and a Cancel rather than an indeterminate spinner. If your source is one mixed track, it says so: a levelled mix leaves only 4 to 5 dB between typical and peak, less than a reaction needs to clear, and detection wants the mic.

Which track is your voice

Import a multi-track OBS recording and the app asks once, right then, which track is you. It lists every track with its codec, channels and rate, the title and language tags OBS wrote, and a short audition waveform to listen against. When detection is confident it simply picks, transcribes that track, and tells you on the rail card with a Change link, so the transcript and the waveform can never end up reading different tracks. Separately, Project Settings gains an Export audio track section for multi-track sources, deciding which stream every export writes and which one the loudness chip measures. That is deliberately a different choice from the voice track: on a typical OBS layout, track 1 is the mix your viewers should hear and track 2 is the mic transcription should read.

Correct every match at once

Correct mode gains a Correct-all card: a find and a replace field, a live list of every match with click to jump and a checkbox to exclude any you want left alone, and one button that corrects all the rest as a single undo step. Phrases spanning several words work, and a replacement with a different word count is aligned deterministically rather than guessed at. Alongside it, each project now keeps a term list: comma separated product names and brand spellings that bias the recognizer at every transcribe, so the name you say forty times stops coming back forty different ways. The Parakeet family honors the list; the whisper models ignore it without complaint.

Import the markers you already made

If you hit a marker hotkey while recording, that work now comes into the editor. Import chapter markers reads the source’s own embedded chapters, the output of the OBS 30.2+ hybrid MP4 and MOV marker hotkey, whose timestamps are recording relative by construction, so there is no offset to work out. From file (CSV)… reads Twitch marker exports and the CutCue and Premiere marker family, recognized by content rather than by file extension, with a signed seconds offset for the usual case where the stream started before the recording did. Every marker arrives as a label, a point marker becoming a 30 second window with the timestamp two thirds in, all as one undo step with the imported and skipped counts reported. They are proposals: nothing is struck, kept, or clipped on your behalf.

The waveform draws itself while it scans

A whole-file waveform scan takes about 11 seconds on stream-length footage, and until now that was 11 seconds of uniform skeleton. The core hands the peaks over as they close, so real bars fill in over the part that has been decoded and the skeleton covers only what is left. The waveform is its own progress indicator, which is how the audio tools you already use do it. The same groundwork gives every heavy read-only scan a percentage and a Cancel, and a scan you cancel stops decoding rather than running on unseen.

Transcription memory stops tracking recording length

Transcription used to hold the entire decoded audio in memory for the length of the run, roughly 230 MB per hour of source. It now streams the decode to a temporary file and reads it back through a memory map, so the engines see byte-identical audio while the resident cost stops scaling with duration. On a three hour recording that is 722.8 MB of peak memory down to 40.5 MB, with an identical transcript. Every engine benefits, since they all sit on the same feed.

Variable frame rate exports stay in sync

Matroska reports the frame rate the encoder was nominally set to, not the rate it actually delivered, and the exporter believed it. On variable frame rate captures from OBS and ShadowPlay, that compressed the video against the audio: a 30 second test file came out with 9.9 seconds of video against 29.0 seconds of audio, and a real recording dropping around 1% of its frames drifted about 70 seconds across a two hour VOD. Frames are now timed from their own presentation intervals instead of a frame counter, on both the CPU and the zero-copy export paths. Constant frame rate sources export bit for bit as before.

Also in this release

  • Panes are separated by a single hairline with a 12px drag zone on it, the reader’s line length widens with the window (roughly 50, then 60, then 71 characters), and the app opens at 1152×720.
  • Undo and redo moved out of the top bar and into the transcript toolbar, alongside the edit history they belong to. The keyboard shortcuts are unchanged.
  • The re-transcribe model picker goes wide rather than tall, so its buttons stay on screen on a short window.
  • A model file left over from an earlier catalog pin is caught by its size and re-fetched, instead of being used as though it still matched.
  • Matroska opens cap their analysis budget in the audio decode lane, the one scrubbing, zooming, and auditioning pay over and over.
  • The in-app Help dropdown is gone: macOS and Windows both have a native Help menu, and it carries the same links.
  • A transitive PostCSS dependency is bumped past a published advisory.

v0.2.4

TimelineWaveforms

The timeline opens on the whole recording, and the waveform underneath it stops fighting you.

The whole file, on open

Opening the timeline used to drop you into an arbitrary window of the recording. It now starts zoomed to the entire file, so the first thing you see is the shape of the whole thing: where the talking is, where the silences fall, and how much of it there is.

A waveform that keeps up

Four separate defects were making the waveform feel unreliable at speed, and all four are gone. The refined, high-resolution pass arrives sooner and lands with far less of a visual jolt. The mini strip draws its bars at the width it actually has rather than a width it assumed, and its playhead now lands on the waveform it names. Zoom refinement, which was quietly unable to fire at all, fires.

v0.2.3

SupportPerformance

Getting help without leaving the app, plus two decode paths that stopped doing redundant work.

Ask for help from inside the editor

Reporting a bug, sending feedback, and asking a question no longer mean hunting for a link. All three live in the app now, pointed at the right place.

Less work per waveform fetch

Two performance changes target the same hot path. The resolved output rate is memoized, so a windowed fetch opens the container once instead of once per call. And a lazily parsed Matroska index gets nudged into shape rather than walked end to end, which matters most on the long MKV recordings where that walk was most expensive.

v0.2.2

WaveformsRotationFixes

Waveforms stop loading whole files into memory, and rotated footage finally arrives the right way up.

Waveforms that never load the whole file

Waveform decoding is streaming now: memory scales with the number of buckets drawn rather than the length of the recording, and no whole file of PCM is ever held. Peaks are fetched in windows sized to the current zoom, so the cost of drawing tracks what you are actually looking at. Loudness measurement got the same treatment and streams rather than materializing the whole timeline. While audio is still being read, the strip says so instead of drawing a convincing line of silence.

Rotated footage, the right way up

Video shot on a phone carries a display matrix telling players which way to turn it. Decoded pixels honor that matrix now, and the probe reports display dimensions rather than the stored ones, so a portrait recording stops arriving on its side and stops exporting that way too.

Also in this release

  • An implausibly sparse container index is distrusted rather than believed.
  • A bulk filler strike rides one core compound instead of a covering band.
  • An empty project looks empty: a real waveform, no stale card.
  • Every modal answers Escape and Enter.
  • One reference-counted core://event subscription replaces the per-consumer ones.
  • libav’s log level is clamped, so the decoder stops flooding stderr.
  • Windows code signing moved from PFX and signtool to SSL.com eSigner.
  • Each Sentry release is tagged with the shipped tag rather than the frozen crate version.

v0.2.1

Windows

Windows joins the release channel.

Windows builds ship

The release channel publishes Windows alongside macOS, with a public-repo publish step and a windows.json updater manifest so installed copies can update themselves.

An encode ladder that degrades gracefully

Windows export runs a ladder of encoders: vendor hardware encoders first, a Media Foundation fallback underneath, and zero-copy wiring through the middle. A false positive on AMD’s AMF encoder used to dead-end the export instead of stepping down to h264_mf. It steps down now.

Also in this release

  • Hardware accelerated FFmpeg and CrispASR are built for Windows in CI, with libass provisioned and bundle DLLs mapped flat next to the executable.
  • Native File menu enable and disable calls are marshalled onto the main thread, and the menu is rebuilt on project open and close rather than mutated in place.
  • Resumed model downloads recover from an HTTP 416 instead of giving up.
  • Projects store absolute source paths when no relative form exists.
  • The Sentry DSN is wired into release CI, and download failures are captured.

v0.2.0

ShortsClipsFramingCaptionsLoudness

The whole short-form toolkit arrives at once: clips, vertical framing, burned-in captions, and exports conformed to what the platforms actually want. This is the big one.

Clips, and the vertical frame they go out in

Select a range in the transcript, press ⌘⇧K, and it becomes a clip, highlighted in the text with its own rail card and a per-platform duration badge. Whenever a clip exists, Export turns into a split button whose primary action exports that clip: its own MP4 with a re-timed SRT beside it, from one save dialog. Frame presets sit alongside it (Original, Vertical 9:16, Square 1:1) with a crop overlay carrying platform framing guides, a safe box, and a thirds grid, so you can see what survives the reframe before you render. A short clip pre-arms 9:16 unless you have already picked a preset yourself.

Framing that finds the face

Arming a crop preset runs a one-shot, on-device face detection that pre-fills the crop position, centering the speaker in the safe box, with an Auto-frame button to run it again. A position you set by hand is never re-clobbered without that explicit re-run. A Detections toggle under the preview draws what the detector actually sees on the frame you are paused at, confidence score included, so a poor suggestion is explainable rather than mysterious. Every degraded outcome falls back silently to the centered crop.

Captions burned into the picture

Burned-in captions become a project setting with two states, Off and On, and no ambiguous Auto. When they are on, the caption stream is composited into the exported frames, and the preview pump composites the same stream from the same setting, so what you watch is what you ship. Placement is decided once per export, choosing the safe-box position that keeps the words readable without covering the detected face, and never moving mid-video.

Exports conformed to what the platform wants

Every video export sends an encode spec through a single seam, and the core conforms the encode to the platform-recommended upload profile: High Profile, closed GOP, CABAC, BT.709 tagged and converted, HDR rejected up front rather than at the end. Short-form adds the ingest overlay on top. The Deliverable Mode dialog reads source against target row by row, so you see "Source: 3840×2160, will scale to 1920×1080" before you commit rather than after. Frame rate can be clamped into the 23 to 60 band, and export loudness gains a video flavor: Off, −14 LUFS Social under a −1.5 dBTP ceiling, or Custom.

Say what you are making, once

After the first import into a new project, a chooser asks what you are making: Long-form, Short-form, or Podcast. Confirming applies the matching presets in one shot (Short-form arms 9:16, turns burn-in on, and sets a three-minute target) and then gets out of the way. It is never re-imposed, and anything you change afterwards stays changed. Skip it and nothing is applied at all. The chosen mode also shapes which panels are on screen, and the native View menu puts any of them back for good.

Two read-outs that keep you honest

An optional target runtime shows a live under or over read-out against the edited timeline, and inline timecode markers appear at each thirty-second boundary of the post-cut timeline, reshuffling as you strike and restore material. Both answer the same question: how long is the thing you are actually making. Separately, if the measured loudness of your source sits far from the target while normalization is Off, the chooser says so, and says in which direction.

Also in this release

  • Audio-only exports gain a conformance clamp: sample rate snaps into the 44.1 or 48 kHz family, and MP3 bitrate lands in a mono or stereo band resolved from the source’s own probed bitrate. A source without a probed bitrate passes through unclamped rather than guessing at one.
  • tract-onnx, tract-nnef, and time were bumped past Dependabot advisories.

v0.1.13

CaptionsTimelinePerformance

Captions learn where a sentence ends.

Cues break where the sentence breaks

Caption cues now group at punctuation instead of at arbitrary word counts. A cue holds a clause or a sentence, which is the difference between subtitles that read and subtitles that merely appear, and it costs you nothing: the grouping comes from the transcript you already have.

Also in this release

  • The timeline playhead and scrub map through full-axis kept segments.
  • absorb_nonspeech_gaps and IntervalSet::union are linear rather than quadratic.
  • The scrub-preview decode is bounded and cancellable.

v0.1.12

Silence

Silence becomes as easy to cut as an "um".

Silence you can delete

Over-long pauses are detected alongside filler words and struck in exactly the same way: a bulk Remove silences action for the whole recording, or gap chips in the reader to step through them one at a time. Playback and export skip a struck silence the way they skip a struck word, and one click restores it. Dead air was always the easiest thing to cut and the most tedious to find.

Also in this release

  • A failed silence strike is no longer masked as a success, and the timeline clip’s extent is restored when the project is reopened.

v0.1.11

TranscriptionTranscript

A transcript release: real pauses come back, and a struck word can be un-struck.

Parakeet reports true pauses again

Word timestamps coming out of Parakeet had been swallowing genuine pauses, which left the transcript reading as one unbroken stream and gave silence detection nothing to find. The timestamps are honest again, and everything downstream of them reads correctly.

Backspace restores what it struck

Pressing Backspace or Delete on a word that is already struck restores it, rather than doing nothing and leaving you to hunt for the undo.

Also in this release

  • Identify speakers defaults to unchecked.
  • The supporting pane scrolls instead of compressing its cards.

v0.1.10

ImportProjects

Import and project state stop surprising you.

Re-importing behaves

Importing the identical file a second time is now a true no-op rather than something that quietly duplicates state. Importing a different file resets the annotations layer instead of leaving markers pointed at media that is gone, and the project frame rate follows whichever import is actually active.

Also in this release

  • The primary source resolves through the active timeline clip rather than the first row in the sources table.

v0.1.9

Fix

A single scoping fix: transcript reads are bounded to the project’s active source.

v0.1.8

CaptionsChaptersLoudnessSpeakers

Captions, chapters, and loudness normalization arrive together, and playback gets a great deal quieter under the hood.

Captions follow the edit everywhere it goes

A caption export operation with its own File menu entry writes subtitles from the corrected, post-cut transcript. Those captions are also baked into the FCPXML timeline export and into the MP4, so the file you hand off already carries them.

Chapters

Chapter markers live in the Annotations layer, with a transcript UI for pinning them to words. Export produces a YouTube-ready block, puts it on the clipboard, and writes FCPXML markers.

Loudness and speakers

Audio export presets normalize loudness to a target. Diarization improved on two fronts: VAD-turn diarization is wired onto the whisper ladder, and pyannote with an embedder handles re-identification inside diarize_words, so a voice keeps its identity across a whole file.

Playback that stops thrashing

The play path streams audio through a bounded ring buffer, a keyframe index is cached with a container fast path behind it plus warm and watchdog awareness, and import hashing runs in parallel. Scrub-seek decoding anchors at the keyframe nearest the requested time instead of somewhere adrift of it.

Also in this release

  • The supporting side panel is resizable and collapsible.
  • A transcript toolbar (broom and fade) and a unified toast replace the earlier ad hoc surfaces.
  • FCPXML asset start anchors at the media’s probed timecode rather than an assumed one-hour origin.
  • The macOS minimum system version drops to 11.0 to match the documentation.

v0.1.7

ProjectsTranscriptionDesign

Projects become documents, whisper.cpp is gone, and the app puts on its warm-editorial clothes.

A project is a file, and there is a place to keep them

A default library at ~/Documents/Scenaristo holds your projects. The File menu gains New, Open, Close, and Save As with native dialogs, backed by an app-level state store and Recent Projects. Save As and Duplicate Project are implemented with SQLite’s VACUUM INTO, so a copy is a genuinely clean copy. Importing media anchors an untitled project into the library rather than leaving the work homeless, a home surface lists Recents, and a welcome project is seeded on first launch.

One transcription ladder

whisper.cpp and whisper-rs are removed, and the whisper ladder runs through CrispASR alongside Parakeet. The whisper catalog is trimmed to large-v3 and large-v3-turbo, and long Parakeet transcriptions report chunked progress live instead of going quiet.

The warm-editorial refresh

The interface takes on the design the app wears today. Alongside it: caret keyboard navigation in the transcript, so the whole document is reachable without the mouse, and a toggle between a subtle and a marked style for the filler words the app highlights, because how loudly the app should point at your "um"s turns out to be a matter of taste.

Also in this release

  • The project WAL is checkpointed on shutdown and on idle, and the -wal and -shm files are actually removed on a graceful exit.
  • Project lifecycle operations are non-blocking, and Seek, open_frame_channel, and ExportVideo decode off the dispatcher mutex.
  • Playback PCM is resampled when the output device rejects the source rate, and silent play failures surface as a toast.
  • A second media import replaces the prior clip instead of appending to it.
  • The Import Media dialog no longer reopens in a loop after a cancel.
  • Preview and play run to an empirical decode budget, resize VideoToolbox frames on the GPU, and reseek keyframe-aware for large forward jumps.

v0.1.6

PlaybackReliability

Playback gets a new engine, and the app starts telling us when it crashes.

Play that does not get slower as the file goes on

Preview playback moves onto a canvas-direct path fed by a sequential decoder. The path it replaces decoded every frame independently, so the further into a recording you played, the more work each frame cost. That cost is gone: playing at the ninety-minute mark is as cheap as playing at the start.

Crashes and hangs, reported

Crashes and hangs are captured through Sentry, with an IPC hang watchdog for the cases where nothing crashes and nothing moves either.

Also in this release

  • Spanish transcribed as English is fixed, with selectable Canary and Cohere engines, ECAPA automatic language detection, and an overlap-tiling correction.
  • Backspace and Delete in the transcript reader are caret-based.
  • Changelog generation is automated: a draft workflow, a release body, and a pre-push guard.

v0.1.5

Updates

The app tells you when there is a new version, and installs it for you.

In-app updates

A notification appears when a new version is available, and you can install it without leaving the app. Auto-update is disabled under tauri:dev, so development stops raising spurious update prompts.

v0.1.4

macOS

The crash on signed and hardened macOS builds when opening a file dialog is fixed.

The file dialog opens

The crash traced back to a missing NSOpenPanel entitlement. That entitlement is declared now, along with the folder-access usage strings macOS requires before it will let the app ask for anything.

v0.1.3

Updates

The foundation for in-app updates.

Signed updates, wired in

The Tauri updater plugin is wired in, so the app can receive and apply signed updates from the release endpoint. The visible half of this arrives two releases later, in 0.1.5.

v0.1.2

ProjectsmacOS

Schema migrations stop being scary, and the signed app launches.

A backup before every migration

When the app needs to migrate an existing project’s schema, it takes a backup first, so nothing is lost if the migration goes wrong.

Also in this release

  • The signed macOS .app failed to launch because the CrispASR and FFmpeg dylibs were not bundled. They are now.

v0.1.1

Playback

The playhead reports the position you are actually hearing.

Sync you can trust

The playback clock subtracts audio output latency and reports the heard position, so the highlighted word in the transcript matches the word coming out of your speakers.

v0.1.0

First releasemacOS

The first distributable release: a signed and notarized macOS DMG.

Transcription and speakers

CrispASR with Parakeet is the default transcription engine, replacing whisper.cpp. Speaker diarization produces stable identities across a file through CrispASR’s global diarizer, and those speakers persist between sessions and can be renamed. Transcription is non-blocking, with live progress events, and hardware-acceleration detection is shared across the backends.

Playback and preview

Playback runs in real time on an audio clock through cpal, with video as the master clock and frames painted while you scrub. Preview rides a program stream: a loopback transport, a fragmented-MP4 MSE surface inside WKWebView, and cooperative render cancellation.

Export

MP4 export at source resolution honors every cut. FCPXML 1.11 serializes the edit for Final Cut, with an Export Timeline menu item to reach it. On Apple silicon, an opt-in zero-copy VideoToolbox path takes decode straight into encode.

The editing surface

Drag and drop import with a native drop target and a visual overlay, a Script view laid out as a US Letter manuscript with screenplay character cues, word corrections dispatched to core as you type, and a splash screen that shows the hardware-acceleration checklist while the app starts.

Also in this release

  • Seek extraction starts from the real keyframe landing, not the requested anchor.
  • Core resolves the program-stream source from the timeline instead of expecting the UI to pass a raw source id.
  • Scrubbing is latest-wins, through seek-sequence correlation and coalescing in core.
  • The project frame rate is set from the imported media, which fixes export and preview drift.
  • Contiguous deleted-word cuts merge, absorbing the inter-word silence cleanly.
  • Import hash progress and the download indeterminate bar render correctly.
Keep it coming

Download the latest build,
then tell us what broke.

The app updates itself, so you only have to do this once. Bug reports and feature requests reach a human: there is a link inside the app, or write to us directly.

Download freeSend feedback

Free today · no account required