OpenDub

Local AI video dubbing. Your voices, your timeline, your control.

▶ Watch the demo

Download · Getting started · Features · Documentation · Report an issue

OpenDub is a desktop studio for translating and dubbing videos with local AI. Separate dialogue from music and effects, identify speakers, translate their lines, and generate new speech using voice and emotion references. Review everything against the video before exporting.

Run the complete workflow with Auto dub, or take control of individual segments in Studio and Segments. Once the models are downloaded, the default dubbing workflow runs on your computer without a cloud inference service.

Early preview: OpenDub is under active development. Windows is the most extensively exercised platform; Linux and macOS packages are available for testing. This README describes the current main branch. Consult the release notes for the build date and source revision of each package.

Download

Platform	0.1.0 preview	Start the app
Windows x64	Download ZIP	Extract into a writable folder and open OpenDub.exe.
Linux x64	Download AppImage	Make it executable, then launch it. Targets Ubuntu 24.04 or a compatible newer desktop.
macOS Apple Silicon	Download DMG	Open the DMG and copy OpenDub.app to Applications.

Packages include the application runtime, inference tools, FFmpeg and VLC. No separate .NET SDK, Python environment or compiler is needed to run a package. AI model weights are downloaded separately. Checksums are available on the release page.

Windows packages include CPU and NVIDIA CUDA runtimes; CUDA requires a compatible NVIDIA driver. Linux packages include CPU, CUDA and Vulkan audio runtimes, plus CUDA translation. macOS packages include Metal audio and translation runtimes. The macOS preview is unsigned and not notarized, so Gatekeeper may block it.

System requirements

These are practical recommendations for OpenDub's default GGUF models, not a claim that every workload has a fixed memory requirement. Longer speech, source separation, other running applications and GPU drivers affect peak memory.

Resource	Practical starting point	Recommended for regular dubbing
Operating system	Windows x64; Ubuntu 24.04 or compatible Linux x64; macOS 14+ on Apple Silicon	Current OS and compatible GPU drivers
CPU	Modern 64-bit processor, 4 cores	8 or more cores
System RAM (Windows/Linux)	16 GB	32 GB or more
NVIDIA GPU for IndexTTS2.5 Q8	8 GB VRAM as the recommended minimum; the development machine uses an RTX 3070 8 GB	12 GB+ VRAM for more headroom
Apple Silicon memory	16 GB unified memory as a starting recommendation, not a validated lower bound	24–32 GB+ unified memory
Storage	At least 20 GB free for app, models and basic working space	SSD with additional space for source videos, stems, temporary WAVs and exports
Network	Required for the initial model downloads (about 8 GiB)	Not needed for the default local dubbing workflow after setup

8 GB VRAM is not a universal hard cutoff. CPU inference is available without a dedicated GPU, but can be substantially slower. Smaller GPUs may run some tasks, yet are not the recommended configuration for the complete IndexTTS workflow. Apple unified memory is shared by the system and GPU and is not equivalent to dedicated VRAM. Close other GPU-heavy applications before generation.

GPU selection: Advanced Settings provides Auto / CPU and explicit accelerator choices. Auto checks actual devices and falls back to CPU when no supported GPU is available. An explicitly selected but unavailable backend produces an actionable error instead of silently claiming GPU acceleration. NVIDIA Linux CUDA binaries target Turing or newer GPUs (compute capability 7.5+); Vulkan provides an additional audio inference route for supported NVIDIA/AMD/Intel drivers. macOS uses the bundled Metal runtime on Apple Silicon. Driver support and model support still matter.

See the GPU validation notes for measured memory, runtime checks and the limits of testing on WSL and hosted macOS.

Getting started

1. Prepare the models. On first launch, the setup assistant downloads six pinned models sequentially, shows progress and verifies SHA-256 checksums. Allow approximately 8 GiB for weights, plus space for the app, media and generated audio. Cancelled downloads can resume; an existing model folder can be reused.
2. Import a video. Choose the dialogue audio track. Enable separation when speech, music and effects are mixed together, or use an available background track.
3. Review the dialogue. Inspect transcription, speaker assignments and timings. Choose the output language and each speaker's voice reference.
4. Translate and generate speech. Process everything or a selection such as 3-5,9,12-15. Preview the result and regenerate any line that needs work.
5. Export. Save a dubbed video, an audio file or an editing package.

Prefer a single flow? Auto dub runs analysis, translation, speech generation, duration fitting and export, saving checkpoints along the way. With a project open, it resumes that project. Import another video to start a new job.

Features

A studio built around dialogue

• Video and waveform timeline: synchronized playback, a draggable playhead, zoom, and automatic selection of the current segment.
• Independent audio controls: mute and adjust the original voice, dubbed speech, and available music/effects tracks.
• Side-by-side text editing: original and translated dialogue in the Segments workspace, with video preview and segment actions.
• Precise segment tools: create, split, delete or merge neighbouring segments; drag waveform boundaries and optionally transcribe newly selected audio.
• Visible occupied ranges: existing segments appear in the Create editor, and overlapping new selections are blocked. Adjacent segments may share a boundary.
• Overlap review: compare original and dubbed waveforms, adjust timings or text, merge and regenerate, trim excess audio, or keep the overlap.
• Recent projects and activity logs: reopen the last ten projects and follow progress in resizable logs or operation dialogs.

Local AI with editable results

• Speaker-specific voice references: select a detected segment, adjust its reference range, or provide an external recording.
• Voice and emotion conditioning: use the original speech as an emotion reference while choosing the speaker reference for voice identity.
• Local translation choices: bundled llama.cpp with Qwen3 4B Q4_K_M, or models available through a local Ollama installation.
• Editable translation prompts: add guidance such as regional Spanish or a preferred style before starting a translation pass.
• Selective processing: choose all segments or numbered ranges, with an option to overwrite existing translations or speech.
• Duration fitting: bounded synthesis retries and permitted time stretching help speech fit its slot. Exact timing and voice similarity are not guaranteed.
• Simple and advanced settings: everyday preferences up front; model, runtime, backend and diagnostic controls available in Advanced.

The current IndexTTS2.5 integration supports dubbed speech in Chinese (zh), English (en), Japanese (ja), Spanish (es) and Arabic (ar). Transcription has its own language support. The interface defaults to English and uses JSON localization files, with Spanish included.

Fast iteration without rebuilding the whole soundtrack

During playback, OpenDub mixes individual segment WAVs in small PCM blocks alongside the muted source video. Regenerating one segment does not rebuild a movie-length preview WAV. A full soundtrack is rendered for export.

Text edits are saved automatically after a short pause. Changing a translation or speaker requires new speech; moving a segment keeps its audio, while changing its length invalidates duration fitting. Following segments never shift automatically. See the audio segment lifecycle for details.

Export

Output	What you get
Recommended MP4	H.264 video, AAC audio and optional selectable subtitles.
Preserve video quality (MKV)	Original video stream copied for full-length exports.
Audio only	MP3, WAV, FLAC, M4A or OGG. MP3 defaults to 192 kb/s stereo; WAV uses 24-bit PCM.
Editing package	Dialogue WAV, available background WAV, mixed WAV and SRT subtitles.

Export the whole project or a time range, using start/end times such as 00:00:00 to 00:10:05. Video, audio and subtitles are cropped together and start at zero in the result. Video range exports are re-encoded for accurate cuts, including MKV.

Exports use the timeline's mute and volume settings. Video profiles can retain original audio as alternative tracks. Progress covers validation, mixing, encoding and publication into an empty destination folder.

Duration mismatches are allowed by default in Export and Auto dub. Speech keeps its start position and may overlap the next segment within the exported range; stricter timing remains optional. Missing, stale or externally modified audio still requires attention.

After export, two highlighted links below the project name open the exported file and its folder. The destination is remembered when reopening a project; older exports can be discovered in the project's standard output folders.

AI pipeline

Stage	Default model / tool	Purpose
Audio preparation	FFmpeg	Inspect tracks, extract audio and prepare media.
Optional source separation	BS-RoFormer	Separate dialogue/vocals from background audio.
Transcription	Qwen3-ASR 0.6B	Recognize speech.
Speaker diarization	Nemotron 3 Diarization	Assign speaker turns.
Forced alignment	Qwen3 Forced Aligner 0.6B	Align recognized words with audio timing.
Translation	Qwen3 4B Q4_K_M via llama.cpp	Translate dialogue; local Ollama is also supported.
Speech generation	IndexTTS2.5	Generate speech with voice and emotion references.
Timing and export	OpenDub + FFmpeg	Fit durations, mix tracks and encode output.

Audio models run through audio.cpp. During a generation batch, the engine can keep its model resident across requests; process transitions release model memory. Logs expose runtime/backend selection and available GPU diagnostics. Acceleration depends on the selected runtime, model and hardware.

Separation, transcription and diarization can make mistakes, particularly with trailers, loud effects, music and simultaneous speakers. Review detected dialogue and references before a final render. Separation may leave traces of the original voice or remove some background detail; duration fitting is not lip synchronization.

Projects and portability

• Windows portable mode: portable.mode keeps settings and models in data/, and projects in projects/ by default. The project folder is configurable in Settings.
• Linux and macOS: default to user storage so read-only AppImages and application bundles remain untouched.
• Portable storage on any platform: set OPENDUB_PORTABLE_DIR to a writable folder.
• Self-contained projects: imported source videos are copied into the project by default, with relative paths for internal media. External references and custom runtime paths must remain accessible.

When updating a portable installation, preserve your data/ and projects/ folders. Project directories use the .dubproj extension. Internal solution and namespace names still use DubbingApp / Dubbing for compatibility.

Build from source

Install the .NET SDK specified in global.json, then:

git clone https://github.com/estebanstifli/OpenDub.git
cd OpenDub
dotnet restore DubbingApp.sln --locked-mode
dotnet test DubbingApp.sln -c Release
dotnet run --project src/Dubbing.App

Native tools and models are also needed for real dubbing; see packaging and runtime setup. To build a Windows portable archive with PowerShell 7:

./scripts/package.ps1 -Runtime win-x64 -AudioCuda -Archive

Linux AppImage and macOS DMG packaging use native build hosts. GitHub Actions builds/tests all three targets and produces package artifacts. Default tests do not download model weights. For synthetic media integration tests, set FFMPEG_TESTS=1 and provide DUBBING_FFMPEG and DUBBING_FFPROBE. Real AI and playback tests are opt-in.

Documentation

• Architecture and project format
• Studio, audio tracks and transcript review
• Segment audio lifecycle
• AI engines and localization
• Packaging and 0.1.0 release notes

Feedback

Open an issue for bugs or suggestions. For a reproducible bug, include your app version, operating system, engine/model, selected backend, steps and the relevant log excerpt. Remove private paths or dialogue from anything you share.

License

OpenDub's original application code is released under the MIT License. You may use, modify and redistribute it, including commercially, while preserving its copyright and license notice.

This does not relicense bundled libraries, inference runtimes, model weights or media. Those keep their own terms. In particular, IndexTTS uses the bilibili Model Use License, not MIT. See third-party notices, the included license texts and redistribution notes.

Acknowledgements

OpenDub would not exist without the work of these teams and their communities. Thank you for making local audio and AI tools available:

Project	Contribution to OpenDub	Upstream license / terms
audio.cpp · ShugoAI	Native audio inference and GGUF model conversions	Apache-2.0 for the pinned runtime; model terms are separate.
IndexTTS · bilibili / IndexTeam	Expressive speech generation and voice references	bilibili Model Use License.
Qwen3-ASR and Qwen3 · Qwen team	Transcription, forced alignment and translation	Apache-2.0 for the selected models.
Nemotron 3 Diarization · NVIDIA	Speaker diarization	OpenMDW-1.1.
BS-RoFormer implementation and audio.cpp model packages	Source separation research, implementation and converted weights	Implementation MIT; the ep368 checkpoint's original provenance remains unresolved upstream.
llama.cpp and ggml	Local LLM inference and tensor computation	MIT.
Ollama	Optional local translation server	MIT; separately installed models have their own terms.
FFmpeg	Media inspection, audio processing, mixing and export	LGPL/GPL depending on build; bundled builds can be GPL.
VLC and LibVLCSharp · VideoLAN	Video/audio playback	LGPL-2.1-or-later for the libraries; VLC/plugins include GPL components.
Avalonia	Cross-platform desktop UI	MIT.
.NET and CommunityToolkit.Mvvm	Application runtime and MVVM	MIT, with third-party runtime notices.
Lucide / Feather	Interface icons	ISC / MIT notices retained.

Original authors retain their copyrights and trademarks. Listing a project does not imply its endorsement of OpenDub. Model weights are downloaded separately and are not included in OpenDub release archives.