Tsuzuri

Features

Audio and text never leave the machine, and the transcription and translation engines come with the app.

Feature	What it does
Transcription	whisper.cpp, with optional VAD
Translation	llama.cpp, guided by a glossary
Proofreading	Waveform, comparison, versions
Speakers	Told apart or named, carried into exports
Export	SRT or plain text, bilingual too
Scripts	Your own Macros and export formats

Installation

Download the installer for your platform from Releases.

Platform	Download	Bundled whisper.cpp and llama.cpp
Windows x64	NSIS installer (*-setup.exe) or MSI	Vulkan build
macOS (Apple Silicon)	.dmg	Metal build
Linux x64	.deb or .rpm	Vulkan build

The installer includes ffmpeg as well. Without a Vulkan-capable GPU driver, or to use another build such as CUDA, choose its executable in the app. Tsuzuri is not code-signed, so the system warns the first time it opens.

Verifying downloads

Each release lists every file's checksum in SHA256SUMS. Run these in the download folder:

shasum -a 256 -c SHA256SUMS --ignore-missing   # macOS, Linux
Get-FileHash <file>                            # Windows PowerShell

On Windows, compare the checksum it prints with the file's line in SHA256SUMS.

macOS

macOS blocks the first launch because the app is not notarized. After moving it to Applications, remove the quarantine flag once in Terminal:

xattr -dr com.apple.quarantine /Applications/Tsuzuri.app

Windows

Two things can stop the first launch on Windows:

When	Do
SmartScreen shows "Windows protected your PC"	Choose More info, then Run anyway
The app does not start	Install the Microsoft Edge WebView2 Runtime (built into Windows 11)

Linux

The packages declare the libraries the bundled engines need, such as libgomp and libvulkan, so install them with the package manager to bring those along:

sudo apt install ./Tsuzuri_<version>_amd64.deb      # Debian, Ubuntu
sudo dnf install ./Tsuzuri-<version>-1.x86_64.rpm   # Fedora

Updates

When	What happens
Tsuzuri opens	Checks for a newer version
Any time	Settings → Version and updates → Check for updates
You choose Update	Downloads, verifies, installs, restarts
A task runs	Updating is refused
No check at launch	Turn it off in Version and updates

At launch Tsuzuri checks tsuzuri.aotoki.me for a newer version on your update channel and offers Update in a notification. Updating downloads the same kind of installer you installed from and checks its signature. Only then does it stop the engines and install, and nothing is downloaded until you choose Update. On Linux, installing asks for your password as apt or dnf would.

Update channels

Update channel	Receives
Stable (default)	Released versions
Preview	Every development build, and each release

Choose the channel under Settings → Version and updates. A preview is labelled with the release it builds on and when it was built. Moving from Preview back to Stable waits for the next release, or choose Roll back to stable now to reinstall the current release straight away. The rpm package offers only Stable.

Getting Started

  Open ▾ → a folder ─▶ pick a Resource ─▶ Transcribe ─▶ Translate ─▶ proofread ─▶ Export ▾

Step	Where
Open a folder	Open → Open a folder
Pick a video	The Resource list
Transcribe	Transcribe, then Transcribe
Translate	Translate, pick a language
Proofread	Click a Segment to play it
Export	Export → a format

A folder is a Project: each video or audio file and its subtitles sharing a name are one Resource. Models download on first use; see Models.

Models

Choose models in Settings; a preset below downloads the first time it is used.

Purpose	Format	Presets
Transcription	whisper.cpp GGML (.bin)	Breeze-ASR-25 (Chinese), Whisper large-v3-turbo and large-v3
VAD, when turned on	whisper.cpp GGML (.bin)	Silero v6.2.0
Translation	GGUF	Qwen3-4B-Instruct-2507
Speaker diarization	GGUF	Nemotron-3-Diarization

Model sources

Source	How
A preset	Downloads into the Hugging Face cache
A file on disk	Any file in the slot's format
A Hugging Face repository	Any file it holds for the slot
A gated repository	Uses the hf auth login token
One project	Its own transcription and translation models

The Hugging Face cache is shared with other Hugging Face tools, so a file already there is not downloaded again. A project's own models suit its language, such as a Japanese model for a Japanese project.

Scripts

Ruby scripts add your own Macros and export formats, such as an interview transcript's layout. Put a .rb file in the folder Settings → General → Scripts → Open folder opens.

Example	Adds
merge_runs.rb	A Macro merging a Speaker's runs
tsv_numbered.rb	A TSV export numbering each Speaker's lines

A script reads a copy of the subtitles and returns new Segments or text, which Tsuzuri checks and writes as one undo step. It reaches no file or network. The script API is in docs/design.md chapter 14.

Contributing

How to build, run and test Tsuzuri is in CONTRIBUTING.md.

Sponsor

Where	Link
Portaly	Sponsor Tsuzuri
In the app	Settings → About → Sponsor

Tsuzuri is free and open source; sponsoring keeps its development going.

License

Tsuzuri is licensed under the Apache License 2.0. The licenses of the engines and everything else it ships are in the license notice under Settings → About.