What's new in 1.1.1.

1Use Doza Assist in ChatGPT and Claude

Use Doza Assist in your favorite frontier AI. Doza Assist now connects to ChatGPT as well as Claude. In Assistant access, one click adds Doza Assist to ChatGPT for Mac, and the Claude Desktop extension does the same for Claude. Then ask the assistant you already use about your interviews: it reads your transcripts and selects, searches them, adds selects to your Clips and builds stringouts. Codex, Cursor and VS Code connect too.

You choose what it sees. Only the projects you switch on are visible to an assistant, and your video never leaves your Mac.

The connector is sturdier.

• It sees a running app, and its tools carry clear titles.
• Malformed messages get proper errors, and odd characters never end the session.
• Selects stay inside the media, and stringouts keep the round-trip rule.
• A select an assistant adds while you are saving is kept.

2Fast transcription in 25 languages with Parakeet v3

25 languages now get Parakeet's fast transcription. Before 1.1.1, only English ran on Parakeet, the fast speech engine; every other language went to Whisper, which is much slower. With Parakeet v3, English plus Bulgarian, Croatian, Czech, Danish, Dutch, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian all transcribe on Parakeet, on your Mac, with no Whisper download.

• The model downloads the first time you use one of these languages, with its progress shown.
• The New Project language hint says which engine each language uses.
• Batch imports pick the right engine for each item, including Auto-detect.

No speech lost at the seams. Long files are transcribed in chunks; words at the joins could be doubled or dropped. They aren't any more.

Transcription status in words. Every phase reads as plain words: extracting the clip's audio, downloading the speech model, transcribing. Time remaining counts from when transcription actually starts. A failed download says so. When Whisper had to stand in, the note has its own Transcribe Again button.

The transcript editor keeps every sentence on screen after you change a speaker. Before, later sentences could vanish until you reopened the project. The word popup also has a Delete button now, and a hint tells you the editor is there.

3Tell Doza Assist what you are making

Tell Doza Assist what you are making. A new short prompt at the top of every project: "What is this piece for?" Write one line about the piece ("a 90-second donor film about the new wing") and AI Analysis, Story Builder, the Story Brief and chat all gear their picks toward that goal.

• Story Builder fills a runtime with the material that serves the goal first.
• The line saves with the project and is optional: leave it empty and everything works as before.

4Story Builder

It sees the whole interview. Story Builder picks from a map of the interview's moments. On long interviews, that map used to have holes, so a "piece about her background" could come back about something else because the background section was never on the menu. Doza Assist now cuts the interview into moments itself and the AI only labels them. Every minute of the interview is on the menu, labelling is faster, and nothing is cut short. If labelling can't finish, the AI Analysis tab says so and offers Label again.

• Projects analyzed with an earlier version show a one-line note on the AI Analysis tab with a Rebuild moments button. It takes a few minutes; running a new AI Analysis rebuilds them too.

It follows your brief. When your prompt names a subject ("her background", "the walk underwater"), Story Builder now chooses moments about that subject before higher-scored moments that aren't.

It lands on the length you ask for.

• A story arrives within 10% of the requested runtime, trimmed on sentence ends.
• The open story shows its runtime against its target.
• "Make it 10 minutes long" means ten minutes. "Make the arc 10 minutes" extends the arc. A length chip replaces a typed length rather than adding to it. A shorter resize is refused instead of cutting the story apart.

Extend a story instead of starting over. Ask for more and the open story grows from its current outline. Added material never replays a moment the story already uses, and a chronological arc stays in order.

Clips start and end on whole sentences.

• Soundbites begin and end on sentence edges, including in Hindi, Urdu, Arabic and Amharic, and in transcripts with no punctuation.
• A run-on sentence still trims to the requested length.
• A clip's start never walks back more than 8 seconds.

Rebuild appears only when the story has a prompt, and rebuilds at the story's length. An edit never cuts your saved prompt. The empty state leads with the prompt box.

Story Brief arcs keep their name, note, prompt and runtime when you send them to Story Builder, and they land on sentence edges.

5Clips play the way you cut them

Clips stop on their out point. The player used to stop up to a quarter of a second late. In a fast-talking interview that was enough to hear the first syllable of the next sentence ("…all year. Sec-"). Every clip now stops within a few milliseconds of its out point. This covers the project page, the collection page, the trim sheet, AI Analysis, chat, chat stories, Story Builder and the Story Brief's Play arc.

Clips start on the first sound. In-points land on the first sound of the word, not where the engine put the word's start. AI Analysis items, chat clips and Story Builder cards now play from the same edges a saved clip and an export use, so what you audition is what you get.

Every clip holds its quote.

• An AI clip whose times don't cover its own quoted words is widened onto them.
• A clip whose quote lies elsewhere is moved there.
• Matching is done on the whole quote in order, so a short filler line ("I think so.") can never pull a clip to the wrong place.
• A story clip's text is now exactly the words that play.

Chat clips stay clip-sized. A chat moment longer than 60 seconds is cut back to the last sentence boundary within a minute. Ask for a duration and the cap follows your ask.

Back and forward 10 seconds beside Play, with J and L (never while you are typing in a field). Back 10 seconds near a clip's start replays the lead-in and keeps the clip.

One player at a time. A clip that finished, or one you clicked away from, no longer pauses or stops the next one. A Play arc survives the picture falling back to the proxy mid-play.

6Speakers and speaker names

Your name, on your voice. When the project's Interviewer field holds your name, Doza Assist gives it to the speaker who asks the questions. This includes a stray second label that speaker detection split off from you. It happens even when the AI only came up with "Reporter".

More real names.

• On 16 GB and larger Macs, a naming pass that finds only role labels ("Reporter", "Interviewer") asks once more, right away.
• Naming runs on the model you picked in AI Model settings.
• Naming waits for a speaker-detection re-run to finish instead of naming labels about to be replaced.
• Two different speakers never get the same real name.

Find speaker names. A button in the Speakers panel on the project page and the collection page, shown while any speaker has no name. One click asks again.

Names arrive without a reopen. The page waits for names and shows them when they land. Names that arrive while you are typing in a Speakers box don't clear it. A new detection run drops a naming that was still waiting, and names left over from an old transcript are cleared by Transcribe Again.

No more "how many speakers?" Speaker detection finds the count itself.

7My Style

Your notes come first. The notes you write on a profile now lead every style prompt and outrank what the analyzer learned. They stay exactly as you typed them, and long notes are trimmed rather than dropped.

Set your typical clip length. My Style measures the clip length from your real finished pieces, and you can set it yourself. Story Builder sizes its clips to it.

Importing references is better.

• Pick whole folders of finished pieces.
• FCPXML bundles import from where they are.
• Each reference transcribes in its own language, on the fast engine, one at a time, and only new files are analyzed.
• A piece exported as both a video and an FCPXML imports once.
• Errors stay on screen until you click Done.

Saving notes or a clip length makes no model call. Regenerate re-reads every piece. Renaming a profile no longer re-scores every project. My Style no longer invents details ("Cuts around …") it can't back up.

The style shapes the story, not just the words. Story Builder, collection search and the moment map score how well each moment fits your active profile. Low-scored moments stay on the menu.

8Chat

• A follow-up answers the new question, not the previous one.
• "Note 2–3 moments for each theme" gives you clips for every theme. It used to cap the whole answer at three clips. A count per theme is now read that way, and a plain total ("give me 5 clips") still wins.
• Pull Clips and Build as Story skip the cards you passed on, and the count says so.
• A short cited line keeps its card. A clip timecode written with a colon before the fraction (00:01:48:68) now plays from the right place.

9AI Analysis and the Story Brief

• Soundbites arrive named, like story beats and social clips.
• Time remaining is honest. The estimate starts when the analysis actually starts working, not while it waits for speaker detection. The wait shows as a wait, so a five-minute queue no longer reads as "80 minutes remaining".
• One model size per run. Analysis, naming and labelling share one context size, so the model isn't reloaded halfway through on smaller Macs.
• The Story Brief fits the model. A long interview's brief is budgeted to the model's context instead of being silently cut. Play arc and Play tension play every segment. Moments play from every timecode form the analyzer writes. An arc sent to Story Builder brings its pitch and runtime.

10Import from Final Cut Pro

• Synchronized clips import straight from a Final Cut Pro browser export, one row per clip, alongside multicam clips. A lone clip in an export imports too.
• Only the part of the file a clip plays is transcribed, and the clip keeps the whole file's audio for playback.
• Multicam and synced clips show the camera picture, in step with the sound.
• B-roll and music beside an interview clip are optional and get no row of their own.
• A project that needs a check says "Check needed" in the Projects list, and every held row says why. The length guideline follows the engine, not the language.

11Exports and documents

• Exports keep your clip titles. "Selects → Download" could wipe clip titles; it doesn't any more.
• An export never deletes a select an assistant just added, and an export of every clip first shows clips added a moment ago. Ticked clips stay ticked when a clip arrives or titles come back.
• Final Cut Pro not on this Mac? The FCPXML is saved, revealed in Finder, and the app says so, instead of reporting a failed export.
• A half emoji in a select or a transcript line no longer breaks every export of the project, or Word and Excel documents.
• A windowed clip's document header gives the clip's own length.
• Quote Sheet: the first Save on the event-title form starts the draft. "From Tagged Clips" keeps one quote per tagged clip, at the clip's time, using only that clip's lines.

12Collections

• Collection Story Builder exports one timeline in story order, cut from every interview's files. It handles frame rates and BWF timecode correctly, including 23.976, 25p and 29.97.
• Two stories with the same title get their own timeline names, and a combined Final Cut export never overwrites an earlier one.
• Builds parse reliably, one click builds once, and saves never race.
• The collection player shows a recorder interview's camera picture, skips 10 seconds with J and L, keeps playing when you switch to audio only, and skips a clip that can't load instead of stopping the arc.
• Collection chat keeps the clips it showed while streaming, finds every card's interview, and drops a clip that cites an interview not in the collection.

13Your own AI key (Claude and OpenAI)

• Analysis runs on Claude Opus 5.5 and the brief and chat on Sonnet 5.5, or on GPT-6 Sol / Luna. Prompt caching cuts repeat cost.
• A connection that drops mid-reply is retried once, then says so, instead of failing the whole analysis.
• A refusal is never saved as an answer.

148 GB Macs, memory and models

Stable through a full analysis on 8 GB. On some 8 GB Macs the screen could corrupt after an analysis because one model load went over the GPU limit. The model now fits, loads once, and a failed call says so. 16 GB Macs keep the faster "draft model".

The app repairs its own models. If the AI model's or the speech engine's files go missing from disk, for example after an uninstall and reinstall, Doza Assist notices at start-up and downloads only what is missing. The splash says "Downloading the AI model again…" and your projects are untouched. Before, the model still looked installed and every AI step failed.

• The app no longer lets the model server delete model files on its own, and it cleans up leftover files after a model update.
• If the model folder disappears while the app is running, the model server is restarted cleanly.

The real download size. The AI model download shows its current size (4.6 GB on 8 GB Macs, 6.6 GB on others), and the bar measures the whole download.

15Progress you can follow

• A Working pill shows what is still running: transcription, speaker detection and naming, builds, batch imports, model downloads, My Style imports. Updated · Show tells you when a page has news.
• The import queue keeps going after you leave the page. A saved import can be stopped, an import a quit app left waits for Continue, and files staged during a run are kept.
• About in the app menu shows the running version and build number, and every build now has its own number.

16Free trial

• The trial notice counts only interviews the 2-minute cap cut short.
• A file no longer than the cap is not called a trial excerpt, and a trial-length transcription settles as "Trial-length" with a working Transcribe Again button.
• Speaker detection runs on the trial's audio, on the file's clock.
• The trial notice links to the right place to unlock your edition.

17FxFactory edition

• A license added in FxFactory unlocks the app within a minute, even if you don't switch back to it, and without reinstalling.
• The FxFactory edition carries everything in this release, the same as the Direct download and Studio.

18Fixes

• A double-click in the transcript editor never starts playback.
• XDCAM MXF and MPEG-TS files report their real frame rate.
• The import guardrail measures placed audio on the timeline, not the source file's length, and its check offers Continue anyway.
• A purpose typed just before a reload still saves.
• Transcribe Again that is rolled back puts back the clip's window and camera.
• The camera picture follows the player's speed and ends cleanly.
• A windowed clip starts at its window even when the browser can't play the video.
• The trim sheet keeps Enter and the arrow keys in the field you are typing in, and its audition plays in proxy and audio-only modes.
• Log lines no longer run into each other, and the app no longer leaves a stray lock behind when it quits.
• Wording: "speaker detection" everywhere.

19Studio: Share and Studio plans

Studio is the same Doza Assist with hosted Share built in. Everything in sections 1 to 18 is in Studio; these are the Studio-only changes since 1.1.0.

Share.

• Documents on a share link. A Documents tab gives collaborators the transcript, selects and quote sheet as Word, PDF or Excel. An Include documents switch in the Studio panel controls it.
• Collections share fully.
  • Collection links publish again, and interview audio plays on them.
  • "Collaborators can use their own AI" works for every interview in a collection link.
  • Highlights and selects from a collection sync back to each interview.
  • A Story Brief moment gets its own play link.
• Windowed clips share correctly. A clip cut from a longer file shares that file's audio, states its own length, and opens at the clip, for you and for collaborators.
• Your work and theirs stay separate.
  • A collaborator's select stays theirs after a reload and a sync.
  • A save you make while a sync is running is no longer reverted.
  • Ticked clips stay ticked when a collaborator's select or feedback arrives.
• Revoking a link removes it. Unpublishing deletes the shared files, and a revoked link answers like one that never existed.
• Leaner uploads. Shared audio is always compressed, never a WAV. A project whose source is a WAV can be shared. The first publish from a new Mac needs no extra setup.
• A plain-words privacy note on the Share sheet and on every reviewer page says what is sent: compressed audio, transcript, clips and documents. Your video never uploads.

Studio plans and trial.

• Studio Teams and owner add-on keys unlock Studio and Share, and every Mac takes a seat. The license window says what kind of key you entered.
• Try Share before paying: the Studio free trial includes two share links, and shares its capped playback.
• An ended Studio subscription locks the app at launch with a clear path to renew.

Studio updates and settings.

• Studio updates itself from its own feed, checks the update's identity before installing, and a ready "Install & Restart" banner never resets itself.
• A ChatGPT tab in Studio settings adds Doza Assist to ChatGPT for Mac.
• A newer Studio build updates the Claude Desktop extension, and Studio replaces an older plain 1.1.0 extension.