tmoroney/auto-subs: Local AI Subtitles for Resolve, Premiere & After Effects
Video editors and developers working with post-production pipelines face a persistent friction: generating accurate, styled subtitles without surrendering footage to cloud APIs or committing to recurring subscription costs. Most existing solutions either require uploading sensitive media to third-party servers, lock advanced features behind paywalls, or produce generic SRT files that still require manual integration into editing timelines. For professionals handling client work under NDA, or teams operating in air-gapped environments, these constraints are non-starters.
tmoroney/auto-subs addresses this directly. It's an open-source, on-device subtitle generator that connects natively to DaVinci Resolve, Adobe Premiere Pro, and After Effects — no cloud processing, no subscription, no data leaving the machine. With 3,809 GitHub stars, 247 forks, and an MIT License, it's gained substantial traction among editors who need local AI transcription with speaker-aware styling. This article breaks down what it does, how it works, and where it fits in your toolchain.
What is tmoroney/auto-subs?
tmoroney/auto-subs is a local-first AI subtitle generation application built primarily in TypeScript. Maintained by Tom Moroney, the project delivers on-device transcription through multiple neural model backends — Whisper, Moonshine, Parakeet, SenseVoice, Canary, and Cohere — all running via whisper-rs (Rust bindings for OpenAI's Whisper C++ implementation) or ONNX Runtime. The last commit was July 12, 2026, indicating active maintenance.
The tool occupies a specific niche: it's not merely a command-line transcription utility, nor a simple SRT exporter. It's a bridge between AI speech recognition and professional video editing workflows. The DaVinci Resolve integration runs through a Lua-based server and Fusion macro system; the Adobe integrations use a bundled CEP (Common Extensibility Platform) extension that communicates directly with Premiere Pro and After Effects.
What distinguishes it from generic Whisper wrappers is the tight coupling with NLE (Non-Linear Editing) timelines. Subtitles aren't just generated — they're styled, speaker-labeled, and pushed back into the editor as native caption tracks or text layers. The speaker diarization feature identifies distinct speakers and enables per-speaker styling, which matters for documentary work, interview editing, and multilingual productions where different voices need visual differentiation.
The project is cross-platform — macOS (Apple Silicon and Intel), Windows (with Vulkan or DirectML acceleration), and Linux (Debian/Ubuntu and Fedora/openSUSE). All models download on-demand through an in-app Model Manager; nothing phones home for transcription.
Key Features
Multiple On-Device Model Families
AutoSubs ships with support for six distinct model architectures, not just Whisper variants. This includes OpenAI's Whisper (via whisper-rs with GGML quantization), Useful Sensors' Moonshine (English and language-specific quantized variants), NVIDIA's Parakeet-TDT-0.6B-v3 (int8 ONNX, 25 European languages plus Russian/Ukrainian), Alibaba's SenseVoice (CJK-focused), NVIDIA's Canary-1B-v2 (with native translation), and Cohere Transcribe (int4 ONNX, highest accuracy for 14 major languages). Model sizes range from 60 MB to 3.1 GB on disk, with RAM requirements from 1 GB to 10 GB.
Speaker Diarization & VAD Beyond raw transcription, AutoSubs integrates speaker identification through a ~40 MB diarization model (user-selectable) and Silero VAD for voice activity detection. This enables automatic speaker labeling and per-speaker formatting — critical for multi-person content.
Native NLE Integration The Resolve integration supports animated caption macros with per-word highlighting, preset style systems, marker-based word timing, and conflict detection. Premiere Pro receives caption tracks directly; After Effects gets SRT entries as text layers. The CEP extension auto-loads when Premiere or After Effects launches.
Built-in Translation Transcription and translation are coupled — generate subtitles in one language, output in another without external tools.
Standalone & CLI Modes Run as a desktop app for simple file-in, subtitles-out workflows, or automate via the CLI for batch operations and pipeline integration.
Cross-Platform Distribution
Native installers for all three major platforms, plus Homebrew support on macOS (brew install --cask auto-subs).
Use Cases
Post-Production Studios with Security Requirements For facilities handling unreleased film, legal depositions, or medical content, cloud transcription violates compliance. AutoSubs keeps all processing local — the README explicitly states "no data leaving your machine." The MIT license permits internal modifications and deployment without legal review of SaaS terms.
Multi-Speaker Interview & Documentary Work Speaker diarization eliminates the manual drudgery of labeling "SPEAKER 1:" / "SPEAKER 2:" in transcripts. Combined with per-speaker styling presets in Resolve, editors can automatically render different speakers in distinct colors, positions, or type treatments — a workflow that previously required manual After Effects composition or expensive dedicated software.
Multilingual Content Operations With 100+ languages and built-in translation, creators producing localized versions can generate base subtitles in the source language, then translate and export without leaving the app. The Cohere and Canary models specifically support native translation, while Parakeet covers European language breadth at high accuracy.
Solo Creators on Variable Hardware The model size/accuracy trade-offs are explicit. A creator on an older laptop can use Moonshine tiny (60 MB, 1 GB RAM) for draft transcription, while a workstation with 10 GB+ RAM can run Whisper large-v3 for final delivery. The Model Manager makes this discoverable without CLI wrestling.
Automation & Pipeline Engineering The CLI mode enables integration with media asset management systems, watch folders, and CI/CD for video pipelines. [INTERNAL_LINK: video-processing-automation]
Installation & Setup
macOS (Apple Silicon)
# Homebrew (recommended)
brew install --cask auto-subs
# Or download the pkg directly
# https://github.com/tmoroney/auto-subs/releases/latest/download/AutoSubs-Mac-ARM.pkg
macOS (Intel)
# Download Intel pkg
# https://github.com/tmoroney/auto-subs/releases/latest/download/AutoSubs-Mac-Intel.pkg
Windows
Download and run the installer:
https://github.com/tmoroney/auto-subs/releases/latest/download/AutoSubs-windows-x86_64.exe
Linux (Debian/Ubuntu)
# Download the .deb package
wget https://github.com/tmoroney/auto-subs/releases/latest/download/AutoSubs-linux-x86_64.deb
# Install with apt
sudo apt install ./AutoSubs-linux-x86_64.deb
Linux (Fedora/openSUSE)
Download the RPM and open with your package manager:
https://github.com/tmoroney/auto-subs/releases/latest/download/AutoSubs-linux-x86_64.rpm
DaVinci Resolve Setup (All Platforms)
Important: The Mac App Store version of DaVinci Resolve is not supported. Download from blackmagicdesign.com instead.
After installing AutoSubs, open Resolve and navigate to Workspace → Scripts → AutoSubs to launch the integration.
Adobe Premiere Pro / After Effects Setup
- Install AutoSubs
- Launch Premiere Pro or After Effects — the CEP extension loads automatically
- Select the Adobe integration from AutoSubs to export timeline audio or import subtitles
Real Code Examples
The README provides limited direct code snippets, which reflects the project's emphasis on GUI and NLE integration over raw API usage. Below are the documented command-line and installation examples with context.
Homebrew Installation
brew install --cask auto-subs
This is the recommended macOS installation path. The --cask flag installs the prebuilt application bundle rather than building from source. After installation, AutoSubs appears in Applications and can be launched normally. The CEP extension for Adobe products is bundled inside the app and registers automatically on first launch of Premiere Pro or After Effects.
Debian/Ubuntu Package Installation
wget https://github.com/tmoroney/auto-subs/releases/latest/download/AutoSubs-linux-x86_64.deb
sudo apt install ./AutoSubs-linux-x86_64.deb
The wget fetches the latest release artifact; apt install handles dependencies. Note the ./ prefix — required for apt to recognize a local file rather than a repository package name. This installs the desktop application, CLI binary, and desktop integration files.
CLI Reference
For command-line usage, the README points to CLI.md rather than embedding examples. This suggests the CLI is substantial enough to warrant separate documentation — typical for tools with extensive flags for model selection, language specification, output formats, and NLE bridge commands. Developers should consult that file for batch operation syntax.
The absence of extensive inline code examples in the README is notable: this is primarily a desktop application with NLE plugins, not a library or framework. Most user interaction is through the GUI or the CEP panels inside Resolve/Premiere/After Effects.
Advanced Usage & Best Practices
Model Selection Strategy Match model to content language and hardware constraints. For English content on limited RAM, Moonshine tiny (60 MB) or Whisper base (150 MB) suffice for drafts. For archival or broadcast delivery where accuracy is paramount, Whisper large-v3 (3.1 GB, 10 GB RAM) or Cohere (2 GB, 4 GB RAM) are justified. The README's accuracy ratings (1-4 stars) provide relative guidance, though absolute WER numbers aren't published — test against your specific audio domain.
Speaker Diarization Tuning The diarization model is user-selectable from the Model Manager. For content with frequent speaker changes (panel discussions, rapid-fire interviews), verify diarization output before applying per-speaker styles — errors here propagate to visual formatting.
Resolve Macro Optimization The v3.5 animated caption macro with per-word highlighting is computationally expensive. For long timelines, consider rendering in segments or using the marker-based word timing for preview, then enabling full animation only for final export.
Version Pinning for Production The project is actively developed (last commit July 2026). For studio deployments, test new releases on non-critical projects before rolling out — the v3.5.1 bug fixes for "Resolve export corrections" and "Linux stability fixes" suggest recent edge cases in those areas.
Comparison with Alternatives
| Tool | Local Processing | NLE Integration | Speaker Diarization | License | Cost Model |
|---|---|---|---|---|---|
| tmoroney/auto-subs | ✅ Full | Resolve, Premiere, After Effects | ✅ Built-in | MIT | Free |
| Whisper (OpenAI) | ✅ Full | ❌ None | ❌ None | MIT | Free |
| Descript | ❌ Cloud | Limited export | ✅ Yes | Proprietary | Subscription |
| Sonix | ❌ Cloud | ❌ None | ✅ Yes | Proprietary | Per-minute |
| Subtitle Edit + Whisper | ✅ Full | ❌ None | ❌ None | GPL-3.0 | Free |
Whisper alone provides the transcription backbone but requires manual SRT import and lacks speaker identification or native NLE styling. Descript offers superior collaborative editing and overdub features, but forces cloud processing and carries ongoing subscription costs. Subtitle Edit is a mature free alternative for Windows-focused subtitle work, but has no AI integration or direct NLE bridging comparable to AutoSubs' Resolve macro system.
AutoSubs' distinctive position is the combination of local AI, speaker-aware output, and deep NLE integration — not any single feature in isolation.
FAQ
Does tmoroney/auto-subs require an internet connection? No. Models download once via the Model Manager; all transcription runs on-device. No data is sent to cloud services.
What's the minimum RAM requirement? 1 GB RAM suffices for tiny models (Whisper tiny, Moonshine tiny). 10 GB RAM is recommended for Whisper large-v3.
Can I use this with the Mac App Store version of DaVinci Resolve? No. Download Resolve directly from Blackmagic Design's website — the Mac App Store build lacks the scripting support AutoSubs requires.
Is there a Windows version with GPU acceleration? Yes. Windows builds support Vulkan and DirectML for model inference acceleration.
What subtitle formats can I export? SRT, plain text, and direct-to-clipboard. Resolve integration pushes styled captions directly into the timeline.
How does speaker diarization handle overlapping speech? The README doesn't specify overlap handling details. Test with your specific audio — diarization quality varies with recording conditions.
Can I contribute or modify the code? Yes. The project is MIT licensed. See CONTRIBUTING.md and the AutoSubs DeepWiki for architecture guidance.
Conclusion
tmoroney/auto-subs solves a specific, well-defined problem: getting AI-transcribed, speaker-labeled, styled subtitles into professional video editing timelines without cloud dependencies or subscription lock-in. It's best suited for post-production professionals, security-conscious organizations, and developers building video automation pipelines who need the transparency and modifiability of open-source tooling.
The 3,809-star traction reflects genuine utility in a workflow gap that commercial tools either ignore or monetize aggressively. The multi-model architecture provides flexibility across hardware and language requirements, while the NLE integrations eliminate the tedious SRT import/export dance.
If your work involves regular subtitle generation in Resolve, Premiere, or After Effects — and you prefer keeping your media local — evaluate tmoroney/auto-subs at https://github.com/tmoroney/auto-subs. The MIT license means you can adapt it to internal workflows without licensing friction, and the active development pace suggests continued refinement of an already capable tool.
