Engines and models

About half of Dogubako’s tools are native and need nothing else. The rest call a mature open-source program (an engine) or use a small local machine-learning model. Everything else keeps working without them.

Who provides the engines

You run	Engines come from
The macOS app	The app itself: built in or downloaded on demand, at pinned versions
The Docker image	The image: every engine is baked in from a pinned Debian snapshot
The CLI, a native server, or a development build	Your system: install what you need (below)

In the macOS app

The app manages its own engines, so there is nothing to install:

Tier	Engines	Where
Built in	FFmpeg and FFprobe, qpdf, Poppler, ExifTool, 7-Zip	Inside Dogubako.app (36 MB of the download)
On demand	ImageMagick (9 MB), Tesseract (10 MB), Ghostscript (20 MB), Real-ESRGAN (51 MB), Pandoc (63 MB), whisper.cpp (~3 MB)	Downloaded into ~/Library/Application Support/Dogubako/engines the first time you need one
System	LibreOffice, Piper, rembg	Not managed yet. Install them yourself and the app finds them

When a tool needs an on-demand engine you don’t have yet, its notice (and Settings → Dependencies) shows a Download button with the size. Downloads are checked against a pinned SHA-256 before they’re unpacked, and nothing is installed system-wide.

Every managed engine is pinned to an exact version, and the app uses only its own copies of them, never one from your PATH or Homebrew, so what runs is what was tested. An app update that bumps an engine downloads the new version when it’s next needed and deletes the old one. Each engine folder carries the license texts of everything in it.

To test a different build of a managed engine, set DOGUBAKO_ENGINES=prefer (use the app’s copies first, then the system) or DOGUBAKO_ENGINES=system (ignore the app’s copies), or point at one binary with DOGUBAKO_<ID>_PATH.

Checking what you have

• In the app: Settings → Dependencies lists every engine with its status (Built in, Downloaded, Not downloaded, Found or Missing), its version, and either a Download button or the install command for your system.
• On the command line: dogubako deps.
• On a server: GET /api/v1/deps.

Detection happens the first time it is needed and is then cached. After you install something, click the refresh button on the Dependencies page or restart the app (POST /api/v1/deps/refresh on a server).

A tool whose engine is missing is never hidden. It shows a notice with the download button or install command, and running it returns a missing_dependency error that names the engine.

Engines

The install commands below are for the CLI and native servers. The macOS app and the Docker image don’t need them.

Engine	Used for	Install
FFmpeg ffmpeg LGPL-2.1+ / GPL-2.0+ depending on build	Video and audio conversion, filtering and encoding 54 tools require it	macOS ffmpegDebian/Ubuntu ffmpegFedora ffmpeg-freeArch ffmpegWindows Gyan.FFmpeg
FFprobe ffprobe LGPL-2.1+ / GPL-2.0+ depending on build	Media stream and metadata inspection 53 tools require it	macOS ffmpegDebian/Ubuntu ffmpegFedora ffmpeg-freeArch ffmpegWindows Gyan.FFmpeg
ImageMagick imagemagick ImageMagick License (Apache-2.0 style)	Extended image formats (HEIC, PSD, ...) and advanced image operations 1 tool requires it	macOS imagemagickDebian/Ubuntu imagemagickFedora ImageMagickArch imagemagickWindows ImageMagick.ImageMagick
libvips vips LGPL-2.1+	Fast, low-memory processing of very large images Optional	macOS vipsDebian/Ubuntu libvips-toolsFedora vips-toolsArch libvips
ExifTool exiftool Artistic-1.0 / GPL-1.0+ (Perl terms)	Read and write metadata for images, video, audio and documents Optional	macOS exiftoolDebian/Ubuntu libimage-exiftool-perlFedora perl-Image-ExifToolArch perl-image-exiftoolWindows OliverBetz.ExifTool
qpdf qpdf Apache-2.0	PDF merging, splitting, rotation, encryption and repair 19 tools require it	macOS qpdfDebian/Ubuntu qpdfFedora qpdfArch qpdfWindows QPDF.QPDF
Poppler utilities poppler GPL-2.0+	PDF text extraction, rendering and image extraction 6 tools require it	macOS popplerDebian/Ubuntu poppler-utilsFedora poppler-utilsArch popplerWindows oschwartz10612.Poppler
Ghostscript ghostscript AGPL-3.0	PDF compression and PostScript processing 1 tool requires it	macOS ghostscriptDebian/Ubuntu ghostscriptFedora ghostscriptArch ghostscriptWindows ArtifexSoftware.GhostScript
Tesseract OCR tesseract Apache-2.0	Optical character recognition for images and scanned PDFs 2 tools require it	macOS tesseractDebian/Ubuntu tesseract-ocrFedora tesseractArch tesseractWindows UB-Mannheim.TesseractOCR
whisper.cpp whisper MIT	Offline speech recognition 3 tools require it	macOS whisper-cppArch whisper.cpp
Piper TTS piper MIT (engine); voices have individual licenses	Offline neural text-to-speech 1 tool requires it	See the project site
rembg rembg MIT (models: Apache-2.0 / various)	Background removal with local U²-Net models 1 tool requires it	See the project site
Real-ESRGAN (ncnn) realesrgan BSD-3-Clause	Local AI image upscaling 1 tool requires it	See the project site
LibreOffice libreoffice MPL-2.0	Office document conversion (DOCX, ODT, XLSX, PPTX → PDF, ...) 2 tools require it	macOS --cask libreofficeDebian/Ubuntu libreoffice-coreFedora libreoffice-coreArch libreoffice-freshWindows TheDocumentFoundation.LibreOffice
Pandoc pandoc GPL-2.0+	Markup document conversion (Markdown, DOCX, HTML, EPUB, ...) 2 tools require it	macOS pandocDebian/Ubuntu pandocFedora pandocArch pandocWindows JohnMacFarlane.Pandoc
7-Zip 7zip LGPL-2.1+ (unRAR restriction applies to RAR code)	7z / RAR archive extraction and creation Optional	macOS sevenzipDebian/Ubuntu 7zipFedora p7zip-pluginsArch 7zipWindows 7zip.7zip
Bubblewrap bwrap LGPL-2.0+	Optional sandbox for external programs (Linux) Optional	Debian/Ubuntu bubblewrapFedora bubblewrapArch bubblewrap

Engines marked optional are not required by any tool as a whole, but some options use them. For example, View Image Metadata reads every tag with ExifTool when it is installed and falls back to its built-in EXIF reader otherwise, and Create Archive needs 7-Zip only for .7z archives. Choosing such an option without the engine gives a missing_dependency error that names it. libvips is still detected but no tool uses it any more, so the macOS app and the Docker image leave it out. bwrap (Bubblewrap) is not used by tools at all; it sandboxes the other engines on Linux servers (see Security).

Pointing at a specific binary

Set DOGUBAKO_<ID>_PATH to use a binary that isn’t on your PATH, for example:

DOGUBAKO_FFMPEG_PATH=/opt/ffmpeg-7/bin/ffmpeg dogubako video compress talk.mov

$env:DOGUBAKO_7ZIP_PATH = "C:\Program Files\7-Zip\7z.exe"
dogubako files archive-extract backup.7z

Or put binaries in Dogubako’s bin folder (see Desktop app), which is searched before your PATH.

Known engine issues

• FFmpeg 9 anlmdn. The non-local-means audio denoiser crashes intermittently with FFmpeg 9 on Apple-silicon Macs. audio.denoise refuses method: anlmdn on FFmpeg 9 and later with an unsupported error and points you to the FFT method, which is not affected.
• Homebrew FFmpeg is built without libvorbis, libass (the subtitles filter), drawtext and animated WebP encoding. Tools that need those report unsupported instead of failing obscurely.
• Tesseract languages. Only the languages you install are available for OCR. On Debian and Ubuntu, add packages such as tesseract-ocr-deu or tesseract-ocr-fra.

Models

Models are downloaded on request, never bundled, and stored in the models folder (DOGUBAKO_MODELS_DIR overrides it). Where upstream publishes a checksum, the download is verified against it; otherwise Dogubako records a local SHA-256 when it first downloads the file.

dogubako models list
dogubako models download whisper-base
dogubako models remove whisper-base

In the app, use Settings → Models. On a server, only admins can download or remove models.

Model	Group	Size	Languages	License
Whisper tiny whisper-tiny Fastest, lowest accuracy (75 MiB)	Speech recognition (whisper.cpp)	74 MiB	multilingual	MIT (OpenAI Whisper weights)
Whisper tiny (English) whisper-tiny.en Fastest English-only model (75 MiB)	Speech recognition (whisper.cpp)	74 MiB	en	MIT (OpenAI Whisper weights)
Whisper base whisper-base Good balance for everyday use (142 MiB)	Speech recognition (whisper.cpp)	141 MiB	multilingual	MIT (OpenAI Whisper weights)
Whisper base (English) whisper-base.en English-only base model (142 MiB)	Speech recognition (whisper.cpp)	141 MiB	en	MIT (OpenAI Whisper weights)
Whisper small whisper-small Higher accuracy, slower (466 MiB)	Speech recognition (whisper.cpp)	465 MiB	multilingual	MIT (OpenAI Whisper weights)
Whisper small (English) whisper-small.en English-only small model (466 MiB)	Speech recognition (whisper.cpp)	465 MiB	en	MIT (OpenAI Whisper weights)
Whisper small + tinydiarize whisper-small.en-tdrz English model that marks speaker turns (465 MiB)	Speech recognition (whisper.cpp)	465 MiB	en	MIT (OpenAI Whisper weights)
Whisper large-v3 turbo (q5) whisper-large-v3-turbo-q5_0 Best accuracy per MiB, quantized (547 MiB)	Speech recognition (whisper.cpp)	547 MiB	multilingual	MIT (OpenAI Whisper weights)
Whisper large-v3 turbo whisper-large-v3-turbo Highest accuracy (1.5 GiB)	Speech recognition (whisper.cpp)	1.5 GiB	multilingual	MIT (OpenAI Whisper weights)
Silero VAD v5.1.2 (ggml) silero-vad Neural voice activity detection used by whisper.cpp --vad (885 KiB)	Voice activity detection	864 KiB	any	MIT
SeetaFace frontal face detector seeta-face Small frontal face detection model used for face blurring (1.2 MiB)	Face detection	1 MiB	—	BSD-2-Clause
Piper voice: English (US) Lessac, medium piper-en_US-lessac-medium Natural US English voice for offline text-to-speech (63 MiB)	Text to speech voices (Piper)	60 MiB	en-US	Blizzard / CC BY-SA (see voice card)
Piper voice: English (UK) Alan, medium piper-en_GB-alan-medium British English voice for offline text-to-speech (63 MiB)	Text to speech voices (Piper)	60 MiB	en-GB	CC BY-SA 4.0 (see voice card)
Piper voice: German Thorsten, medium piper-de_DE-thorsten-medium German voice for offline text-to-speech (63 MiB)	Text to speech voices (Piper)	60 MiB	de-DE	CC0 (see voice card)

Which Whisper model should you pick? whisper-base is a good default on a laptop. whisper-small is noticeably more accurate and still practical on a modern CPU. whisper-large-v3-turbo-q5_0 gives near-best accuracy at a fraction of the full model’s size. The .en models are English-only and a little more accurate for English. whisper-small.en-tdrz adds speaker-turn detection.

Transcription also needs the whisper.cpp engine itself (whisper-cli), and text to speech needs the piper program as well as a voice.