AutoSubs: on-device subtitles that connect to DaVinci Resolve, Premiere and After Effects
On-device subtitle generation that connects directly to DaVinci Resolve, Premiere, and After Effects.
At a glance
- What is it?
- AutoSubs is a local-first transcription app from tmoroney that runs Whisper, Moonshine, Parakeet, SenseVoice and Canary models on your own machine and pushes styled subtitles into three editing hosts. The trade-off is a desktop installer and model downloads instead of a browser tab.
- Who is it for?
- Adopt AutoSubs if you already cut in DaVinci Resolve, Premiere Pro or After Effects and want transcription to stay on your own hardware, with per-speaker styling driven by diarization. Do not adopt it if you need a browser-only tool, a Mac App Store copy of Resolve, or a service that stores transcripts for you.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 1 day ago.
- What is it written in?
- Mainly TypeScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 17, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The problem AutoSubs solves for editors, not for viewers
Subtitle generation is usually a round trip. You export audio, upload it to a service, wait, download an SRT, import it, and re-time it because the service never saw your timeline. AutoSubs collapses that loop: the app runs the speech-to-text model on your machine, then hands the result back to the editor you are already in. The README frames the pitch as local-first. No cloud, no subscription, no data leaving your machine.
The people this is aimed at are specific. If you cut in DaVinci Resolve, Premiere Pro or After Effects and your audio cannot leave the building (client NDAs, unreleased material, medical or legal recordings), a hosted transcription service is a non-starter regardless of price. The same applies if you work offline or on a metered connection.
The README also lists speaker diarization as a feature, which matters more than it sounds. Diarization identifies and labels different speakers in the transcript, enabling per-speaker styling. In practice that means you can style one speaker's captions differently from another's without hand-splitting the track first. That is an editing-time saving, not just a transcription one.
How AutoSubs runs models on-device and pushes them into your timeline
The repository is split into named top-level directories, and they map cleanly onto the three integration paths. AutoSubs-App/ holds the desktop application. Resolve-Integration/ holds the DaVinci Resolve side, and its README is described as covering integration architecture and development. Adobe-Extension/ covers Premiere Pro and After Effects.
The model layer is the interesting part. Whisper runs through whisper-rs on GGML, while Moonshine, Parakeet, SenseVoice and Canary run through ONNX Runtime. That is two runtimes in one app, which explains the size spread across models: moonshine-tiny is listed at 60 MB and 1 GB of RAM, parakeet at 700 MB and 2 GB, large-v3 at 3.1 GB and 10 GB. Models are downloaded on demand from the in-app Model Manager rather than bundled, so the installer stays small and the disk cost arrives later.
The Resolve path is not a plugin in the usual sense. The README tells you to open DaVinci Resolve, then go to Workspace, then Scripts, then AutoSubs. That is Resolve's scripting menu, which means the integration is driven through Resolve's scripting surface rather than a native effect. The Adobe path works differently: a CEP extension loads automatically when Premiere Pro or After Effects is open. In Premiere Pro, subtitles are imported as caption tracks; in After Effects, SRT entries are created as text layers. Those two output shapes are not interchangeable, and the README is explicit about which host produces which.
Installing AutoSubs and running a first transcription
There is no package to install from a language registry. AutoSubs ships as a desktop installer per platform, and the README's download table links a Windows .exe, two macOS .pkg files (Apple Silicon and Intel), a Linux .deb and a Linux .rpm.
On macOS you can also use Homebrew, which is the shortest path if you already manage software that way:
brew install --cask auto-subsOn Debian or Ubuntu, the README gives a two-command sequence that downloads the .deb and installs it with apt:
wget https://github.com/tmoroney/auto-subs/releases/latest/download/AutoSubs-linux-x86_64.deb
sudo apt install ./AutoSubs-linux-x86_64.debFedora and openSUSE users download the .rpm and open it with the system package manager instead.
For a first real use, the standalone path is the one to try before touching an editor. Launch AutoSubs, select an audio or video file, pick your model and language or translation options, then click Transcribe. The README says you can edit speakers and subtitles afterwards and export as SRT, text, or copy to clipboard. Starting standalone isolates model download and transcription from any host integration problem.
Once that works, the Resolve path is: open DaVinci Resolve, go to Workspace, then Scripts, then AutoSubs, select your timeline or audio source and settings, click Transcribe, edit, then send styled subtitles back to Resolve. There is a warning attached to this section that is easy to skim past: the Mac App Store version of Resolve is not supported. Download Resolve from blackmagicdesign.com instead.
For scripted or batch work, the README points at CLI.md, described as a complete command-line reference with examples and troubleshooting. It does not reproduce the flags in the main README, so treat CLI.md as the source rather than guessing at arguments.
Model choice is a RAM decision before it is an accuracy decision
The supported-model tables are the most concrete documentation in the repository, and they are worth reading as a hardware constraint rather than a feature list. Whisper's large-v3 is listed at 3.1 GB on disk and 10 GB of RAM. large-v3-turbo drops to 1.6 GB and 6 GB. medium is 1.5 GB and 5 GB. If you are on a laptop with 16 GB shared between the OS, your NLE and a browser, the top of that table is not realistic while Resolve is open.
The smaller end is genuinely small. moonshine-tiny is 60 MB and 1 GB of RAM for English, and moonshine-base is 200 MB for slightly better English accuracy. SenseVoice is 230 MB and 1 GB, and the README positions it as compact and well-suited to CJK audio, covering Chinese, English, Japanese, Korean and Cantonese. Parakeet is the outlier in coverage: 700 MB, 2 GB of RAM, and 25 European languages plus Russian and Ukrainian, rated at the top accuracy tier alongside large-v3.
Language coverage is not uniform across families, and the README is honest about this. Whisper's .en variants are English-only and slightly more accurate on English audio. Moonshine's language-specific tiny models exist for Arabic, Chinese, Japanese, Korean, Ukrainian and Vietnamese, but the base model is English. If your source language is not in a given family's row, that family is simply not an option for you.
The accuracy column is a relative 1 to 4 rating within AutoSubs, and the README says so directly. It is not a benchmark against other tools and should not be read as one.
Where AutoSubs is the wrong tool
The Mac App Store warning is the sharpest limitation in the README, and it is a hard stop rather than a caveat. If your DaVinci Resolve came from the Mac App Store, the Resolve integration does not work, and the fix is to install Resolve from Blackmagic's own site. That is a real disruption if your studio standardises on the App Store build.
Model downloads are the second constraint. The installer is small because models are fetched on demand from the in-app Model Manager. That is the right default for disk space, but it means the first transcription of a session can stall on a download, and an air-gapped machine needs those files staged some other way. The README describes the Model Manager as the delivery mechanism and does not describe an offline import path.
Diarization is listed as a capability, not a guarantee. The README does not state how many speakers it handles reliably, how it performs on overlapping speech, or what happens with heavy crosstalk. If your material is panel discussions or interviews with frequent interruption, treat speaker labelling as something to verify on your own audio rather than assume.
Finally, this is a desktop application with editor integrations. If your workflow is a browser tab and a downloaded SRT, or if you need transcripts stored server-side for a review team, AutoSubs is solving a different problem than the one you have.
AutoSubs against hosted transcription services
The obvious alternative is a hosted transcription service, and the difference is not accuracy, it is where the audio goes and what you get back. A hosted service accepts an upload from any machine with a browser, returns a transcript, and usually keeps it in an account you can share. AutoSubs requires a local install, downloads multi-gigabyte models, and returns subtitles to a desktop editor. Nothing leaves the machine.
That trade cuts both ways. You gain privacy and offline operation. You lose the zero-install property and the shared web workspace. You also take on the hardware cost: a hosted service's accuracy ceiling is whatever model it runs, while AutoSubs caps at whatever your RAM allows, and the README's own table makes large-v3 a 10 GB commitment.
Within the editor-integration space, the distinction is subtler. Resolve has its own transcription features, and the README does not compare AutoSubs to them. What AutoSubs adds on top of any single host is the same app reaching three of them, plus a CLI path documented in CLI.md for work that does not go through a GUI at all. If you only ever cut in one host and that host already transcribes acceptably, the integration breadth is worth less to you than it is to a shop that moves between Resolve and Adobe tools.
Maintenance, licence and what an upgrade actually costs
The repository is not archived, and the last push was on 2026-09-12. Recent releases are close together: v3.8.0 on 2026-07-21, v3.7.0 on 2026-07-18, and v3.6.2 on 2026-06-09. The two July releases landing three days apart suggests a fast patch cadence around that version, though the release titles alone do not say what changed. release-notes.md sits at the top level of the repository and is the place to look before upgrading.
Upgrade cost is mostly model re-downloads, not code. Because models come from the in-app Model Manager rather than the installer, a new app version does not necessarily invalidate your local model files, but the README does not document whether it does. That is an open question worth answering on your own machine before a large batch job.
The licence is MIT. That is permissive, and it is the same licence family most editors' scripting integrations expect. It says nothing about the models you download through the app, which come from separate upstream projects (OpenAI's Whisper, Useful Sensors' Moonshine, NVIDIA's Parakeet and Canary, Alibaba's SenseVoice) and carry their own terms. Check those separately; the MIT grant on this repository does not extend to them. Nothing here is legal advice.
The repository also carries AGENTS.md, a .agents/ directory and a plans/ directory at the top level, alongside CONTRIBUTING.md and a DeepWiki link the README recommends for understanding the codebase. If you plan to modify the integration rather than just use it, those are the entry points the project itself points at.
Editorial conclusion
Adopt AutoSubs if you already cut in DaVinci Resolve, Premiere Pro or After Effects and want transcription to stay on your own hardware, with per-speaker styling driven by diarization. Do not adopt it if you need a browser-only tool, a Mac App Store copy of Resolve, or a service that stores transcripts for you. Before committing, confirm the Mac App Store warning applies to your Resolve install, check that your machine has the RAM the chosen model needs (large-v3 lists 10 GB), and read CLI.md to see whether the command-line path covers your batch workflow.
Frequently asked questions
How do I install AutoSubs for DaVinci Resolve?
Install the AutoSubs desktop app first, using the platform installer from the release page or brew install --cask auto-subs on macOS. Then open DaVinci Resolve and go to Workspace, Scripts, AutoSubs. The README warns that the Mac App Store version of Resolve is not supported, so use the Blackmagic download instead.
How do I use AutoSubs in DaVinci Resolve?
From Resolve, open Workspace, then Scripts, then AutoSubs, select your timeline or audio source and your settings, and click Transcribe. After editing speakers and subtitles you send the styled subtitles back to Resolve.
How do I use AutoSubs?
In standalone mode, launch AutoSubs, select an audio or video file, pick a model and language or translation options, and click Transcribe. You can then edit speakers and subtitles and export as SRT, text, or copy to clipboard. For scripted use, the README points to CLI.md.
How do I install AutoSubs?
The README lists a Windows .exe, macOS .pkg files for Apple Silicon and Intel, a Linux .deb for Debian and Ubuntu, and a Linux .rpm for Fedora and openSUSE. macOS users can also run brew install --cask auto-subs.
How do I use AutoSubs in DaVinci Resolve?
Open DaVinci Resolve, go to Workspace, then Scripts, then AutoSubs. Select your timeline or audio source and settings, click Transcribe, edit speakers and subtitles as needed, then send the styled subtitles back to Resolve.
How do I auto subtitle?
With AutoSubs, launch the app, select an audio or video file, choose a model and language or translation options, and click Transcribe. The README says you can then edit speakers and subtitles and export as SRT, text, or copy to clipboard.
Community notes