Hushcue

Accurate subtitles from any video, generated privately on your PC.

Hushcue turns the speech in your videos into subtitles with Whisper, running entirely on your own machine. It can save an .srt file, add a subtitle track to an .mkv, or burn the text into an .mp4.

Download: the installer is on the Releases page (Windows 10 1809 or later, x64). Hushcue needs ffmpeg on PATH (winget install Gyan.FFmpeg). On first run it downloads the speech model (about 1.6 GB) and, with an NVIDIA GPU, NVIDIA's cuBLAS (about 630 MB); without a GPU it runs on the CPU. The installer is not code-signed yet, so SmartScreen may warn: More info, Run anyway.

Hushcue is a native Windows app (C# + WinUI 3). It runs OpenAI's Whisper through CTranslate2, called directly from C# through a small C++ bridge, with a C# port of faster-whisper's decoding. What's next: TODO.md.

Finished	Review and edit

Drop videos or audio files on the window (or use Add files), pick the quality and the result, and start. Each file shows its own progress; the bottom bar shows the whole batch and the time left, as does the taskbar button. Files that can't be done are flagged before the run (not media, no sound, or no picture to put subtitles on). The finished screen sums up the batch and can put originals moved to the Recycle Bin back.

Keyboard: Ctrl+O add files, Ctrl+Enter start, Delete remove the selected files, Esc cancel (asks first), arrow keys within the quality and result choices. HUSHCUE_THEME=Light or Dark pins the theme and HUSHCUE_SETUP=1 shows the setup screen as on a fresh PC, and HUSHCUE_NEW_INSTANCE=1 opens a separate window even when Hushcue is already running (all for testing and screenshots).

Layout

Project	What it is
src/Hushcue.Core	Settings, subtitle cleanup, SRT, pipeline, ffmpeg and engine wrappers. No UI.
src/Hushcue.Cli	Console front end, used for benchmarks and scripting.
native/hushcue_ct2	C interface over CTranslate2's Whisper model, built by native/build.ps1.
tools/bench	Scripts that check the C# engine against faster-whisper and write the test fixtures.
eval/lookalike	Spoken test passages for Hushcue.Cli eval: words that sound alike, and names with a vocabulary.
src/Hushcue.App	The WinUI 3 desktop app.
tests/Hushcue.Core.Tests	xUnit tests for Core.

Using it

The sidebar holds the three tools:

• Generate subtitles: drop videos or audio, pick a quality and what you get (an .srt, a subtitle track, or text burned into the picture), and run. Review and edit then goes through what is worth checking: phrases Whisper invents over silence, loops, long gaps, and words it was not sure of that sound like other words (calm / come / cum, their / there, principle / principal, ...). Those come with one-click swaps to the other spellings; "Looks fine" keeps the word. Re-run this part transcribes the selected subtitles again; Try another model does it with another quality's model (Thorough's large-v3 decodes more carefully than the turbo model of Fast and Balanced) and shows what it heard next to what is there, with the words that differ marked, to use or keep. With Preview first on, nothing is saved until you look: each file's subtitles open as a preview in a temporary folder, to Keep (as the result you pick) or Discard. For a bigger video, drag the divider between it and the subtitle list, or hide the list with the button next to the timeline zoom.
• Edit subtitles: open a video with subtitles it already has (talk.srt or talk.en.srt next to talk.mp4 is found by name), check them against the picture, fix text and timing, then save the .srt or put it into a copy of the video. Dropping a video with its .srt, or right-clicking an .srt in Explorer and choosing Edit subtitles with Hushcue, opens it here too. Nothing is transcribed unless you re-run a part.
• MKV to MP4: remove the subtitle tracks from .mkv files and save them as .mp4 without re-encoding.

Settings holds the choices that rarely change (language, vocabulary, audio clean-up, skipping silence, what happens to the files when subtitles go into the video, the theme, automatic updates), and shows where transcription runs on this PC. Translate to English there gives English subtitles for videos in other languages: the language spoken is detected and translated while transcribing (English videos are transcribed as usual). Every quality then uses the large-v3 model, as the turbo model of Fast and Balanced was not trained to translate. About has the version, "Check for updates", the source code and the third-party notices.

Updates: once a day on start, Hushcue asks GitHub whether there is a newer release and offers to install it. The installer is checked against the release's SHA256SUMS.txt, Hushcue closes, updates and opens again. A portable copy is pointed at the release page instead. Turn it off in Settings.

Install

Windows 10 (1809) or 11, x64. From a release:

• Hushcue-<version>-setup.exe installs for your user account (no admin needed), adds a Start menu entry and, if you keep the option ticked, "Generate subtitles with Hushcue" on the right-click menu of video and audio files.
• Hushcue-<version>-win-x64.zip is the same app, portable: unzip anywhere and run Hushcue.exe.

Hushcue needs ffmpeg on PATH: winget install Gyan.FFmpeg. On first use it downloads the speech model (1.6 GB) and, on a PC with an NVIDIA GPU, NVIDIA's cuBLAS (630 MB) into %LOCALAPPDATA%\Hushcue; without a GPU it runs on the CPU, more slowly. Nothing else leaves your PC.

Build

Needs the .NET 10 SDK (pinned in global.json) on Windows 10 1809 or later, x64. The engine bridge also needs Visual Studio 2022 or later Build Tools (C++ workload), CMake, and a Python environment with the ctranslate2==4.8.2 wheel, plus nvidia-cublas-cu12 to run on the GPU during development. By default the scripts look in native\deps\site-packages:

python -m pip install --no-deps --target native/deps/site-packages ctranslate2==4.8.2 nvidia-cublas-cu12

powershell -ExecutionPolicy Bypass -File native/build.ps1
dotnet build Hushcue.slnx
dotnet test Hushcue.slnx
dotnet run --project src/Hushcue.App

On a Mac (Apple Silicon), Core, the CLI and the tests build and run on the CPU; the app itself is Windows only. native/build.sh builds the bridge against the macOS ctranslate2 wheel and puts it, with the CTranslate2 dylib, in native/bin. It needs the Xcode command line tools and CMake. Burning text into a video needs an ffmpeg with libass: Homebrew's ffmpeg has none, ffmpeg-full does.

python3 -m pip install --no-deps --target native/deps/site-packages ctranslate2==4.8.2
native/build.sh
dotnet test tests/Hushcue.Core.Tests
dotnet run --project src/Hushcue.Cli -- job --out out video.mp4

Release

powershell -ExecutionPolicy Bypass -File tools/publish.ps1 -Version 0.2.0 -SmokeTest
powershell -ExecutionPolicy Bypass -File installer/build.ps1 -Version 0.2.0 -Iscc <path to ISCC.exe>

tools/publish.ps1 makes a self-contained folder in artifacts/publish/Hushcue; installer/build.ps1 turns it into an installer with Inno Setup 6. Pushing a v* tag runs the same steps on GitHub (.github/workflows/release.yml) and publishes a release with the installer, a zip and SHA256SUMS.txt, which is what the app's updater looks for. Details, measurements and the ffmpeg and licence choices: docs/shipping.md.

Licence

Hushcue is MIT licensed, by Jatinjit Singh. It ships or downloads other people's work under their own licences: THIRD-PARTY-NOTICES.md.