Building

Run from source

Python 3.14 is required; Windows x64 for the installer.

python -m venv .venv && .venv\Scripts\activate          (Linux: source .venv/bin/activate)
pip install uv
uv pip install torch torchaudio --torch-backend=auto
uv pip install -r tools/requirements-dev.txt
python -m pytest tests                                   (no GPU needed)
python main.py

Never run uv run without --no-sync: it can replace the CUDA build of PyTorch with the CPU one.

Online installer

installer/VoxprintMovieDubber.iss (Inno Setup 6) compiles into installer\Output\VoxprintMovieDubber-Setup.exe (about 2 MB):

ISCC installer\VoxprintMovieDubber.iss

The installer contains the program's own files only. During setup installer/install-runtime.ps1 downloads uv and installs the runtime: Python 3.14 and PyTorch 2.11.0+cu130 pinned by dubber/infra/runtime_lock.json (the Audiobook Builder lock schema; the sibling main lock is still 3.11/cu128 and is not copied). There is no torchaudio wheel for torch 2.14.1 on the cu130 index, so the pin follows the newest official pair: torch 2.11.0+cu130 and torchaudio 2.11.0+cu130. torchcodec is 0.17.0+cu130. nvidia-cublas-cu12 and nvidia-cudnn-cu12 supply the CUDA 12 libraries CTranslate2 needs. The dependencies come from requirements.txt and the pins in installer/runtime-constraints.txt. The runtime lives in %LOCALAPPDATA%\Voxprint\shared\runtimes\py3.14-torch2.11-cu130 (see SHARED-RESOURCES.md). A folder for that version is reused. A different Python or torch line gets its own directory, and a shared runtime is never upgraded in place. <install folder>\runtime is a junction to its env. shared\manifest.json counts the programs using each resource. The last setup step asks VOXPrint MDL (voxprint-mdl enqueue, then status --wait) to download the AI models into the shared models folder. /SKIPMODELS=1 skips that download and is only for CI.

Full (offline) installer

The full installer carries everything the online one downloads, so setup needs no internet (only the optional AI model download does):

powershell -File installer\make-full-payload.ps1 -Out build\payload -Requirements requirements.txt `
    -Constraints installer\runtime-constraints.txt -Lock dubber\infra\runtime_lock.json -Python <python 3.14> -Flavor cu130
ISCC /DFull installer\VoxprintMovieDubber.iss

make-full-payload.ps1 puts into build\payload: uv.exe, the python-build-standalone archive this uv installs for Python 3.14, every wheel of the runtime (pip wheel with the same pins as the online setup, including torch 2.11.0+cu130, torchaudio and torchcodec) and the LGPL ffmpeg archive. The result is VoxprintMovieDubber-Full-Setup.exe plus VoxprintMovieDubber-Full-Setup-N.bin parts, each below 2 GB (the GitHub release asset limit); keep all of them in one folder. During setup install-runtime.ps1 -Payload installs the same runtime key as the online setup from those files, so both installers share one runtime folder. The AI models are not bundled; setup downloads them into the shared models folder. /SKIPMODELS=1 skips that download and is only for CI.

Setup switches: /TORCH=auto|cu130 (PyTorch flavor). /TORCH=cpu installs the CPU build and is only for the CI smoke run on a GPU-less runner; a user install stops instead of offering that build. /SKIPGPUGATE=1 (full installer only) installs its cu130 runtime without the GPU check and is likewise only for CI. /SKIPMODELS=1 skips the model download and is only for CI. /VERYSILENT /SUPPRESSMSGBOXES /DIR=<folder>. cu130 needs NVIDIA driver branch 580 or newer; the hardware gate already requires branch 600.

The setup stops, and installs nothing, when no NVIDIA GPU has compute capability 8.9 or higher (GeForce RTX 40 / Ada or newer) or the driver branch is below 600. The window does the same check with PyTorch before it opens (VOXPRINT_SKIP_GPU_GATE=1 skips that check for tests and CI). The program is installed in C:\Program Files\VOXPrint AI Suite\Movie Dubber. voxprint-movie-dubber.exe opens the window and voxprint-movie-dubber-cli.exe is the console launcher (the same entry point as the headless CLI). The setup puts its shortcuts into the Start menu folder "VOXPrint AI Suite", adds an Apps & features entry and a full uninstaller. The same AppId is kept, so Windows treats a new setup as an upgrade of the old C:\Program Files\VoxprintMovieDubber install: that folder and this app's shortcuts in the old Start menu group "Voxprint" are removed, and user data is kept. Setup writes the per-machine discovery key HKLM\Software\VOXPrint\MovieDubber with InstallPath, Exe and CliExe (full paths of voxprint-movie-dubber.exe and voxprint-movie-dubber-cli.exe), Version, Build (the number from BUILD.json, or the CI stamp when one was present at compile time), IpcName (empty) and ProtocolVersion (DWORD 0). It also writes App Paths for both executables: the default value is the full path and Path is the install directory. An upgrade deletes a dev build's HKLM\Software\VOXPrint AI Suite\Apps\movie-dubber key. Uninstall removes MovieDubber and both App Paths keys, and removes Software\VOXPrint only when that parent is empty. TODO: voxprint-common 0.3.0 suite.registry will own this key. The uninstaller never touches the Audiobook Builder. It drops this program's references in shared\manifest.json and deletes a shared runtime or ffmpeg build only when no program still references it. Projects, logs and reports are removed only when you answer Yes (default: keep).

Shared settings: %LOCALAPPDATA%\Voxprint\state\suite.json (schema 1: ui_language, theme, models_dir, gpu), read and written by both programs (voxprint_common.suite_settings).

Version and codename

BUILD.json at the repository root is the only copy of the build offset and the codename. dubber/appinfo.py reads it. APP_VERSION in that module is the version string (1.0.0-rc). The build number is not stored as a constant.

dubber/appinfo.py asks voxprint_common.buildinfo for the number. On the Build installers workflow that number is GITHUB_RUN_NUMBER plus offset. The workflow writes it, with the codename, into build_info.json. The installed program reads the stamp, so it does not need the git checkout. A missing recorded build is dev, never 0. VOXPRINT_BUILD=<digits> forces that number.

The offset was chosen from the Build installers workflow: run_number 17 on 2026-10-10 (workflow run 38057799983). The next run is 18, and 18 + 983 = 1001.

The window, About, --version, the installer and the GitHub release title show VOXPrint AI Movie Dubber 1.0.0 RC · build N "Codename". The version line itself is 1.0.0 RC · build N "Codename". The splash uses the library label, which has no middle dot. A codename is one Biblical Hebrew word in ASCII transliteration. It describes the state of that build (readiness, a milestone), not a theme of the product. It is shown only when it is not empty, in quotes. The codename of this line is Bochan.

Headless commands for other programs are documented in CLI.md.

GitHub Actions

.github/workflows/build-installer.yml builds both installers on free windows-latest runners (a matrix of online and full), installs each silently (online with the CPU build of PyTorch, full with its own cu130 runtime and no downloads), runs the installed program and the tests, uninstalls it and uploads the installers as artifacts. The workflow runs on a manual dispatch or a v* tag. To build installers on a PR, add the build-installers label or push to build/. Adding that label starts the build; later pushes to the same pull request build again while the label is still there. A push to build/<name> builds that branch. Fork pull requests use the pull_request event, so they do not receive secrets. A v* tag then publishes both on one pre-release; the release job runs only on tags. It checks every .sha256 file and appends a table of file names, sizes and SHA-256 to the release notes (docs/release-notes/<tag>.md).