Kosmos

Download Kosmos for Mac or Windows

Kosmos is a free desktop studio for human-made audiobooks. Import a manuscript, record each chapter or bring audio from another editor, proofread the take against the page, master the sound, and export a delivery-ready book.

The project stays in a folder on your computer. Kosmos has no account, subscription, usage meter, analytics, cloud manuscript upload, AI narrator, or voice cloning. Automation helps with words and delivery specs; a human still makes the final listening and performance decisions.

Lightbox workflow

Every chapter moves through the same three stages:

1. Record — read from a voice-following teleprompter, check the room, record chapter takes, and punch individual lines. You can also import an existing recording.
2. Proofread — compare the take with the manuscript, play each flagged line, leave notes, and turn real mismatches into pickups.
3. Sound — listen to the original, working, and mastered takes; master for ACX or EBU; and inspect loudness, true peak, noise floor, format, and room tone before delivery.

The wider book workspace adds TXT, Markdown, DOCX, EPUB, and text-layer PDF import; chapter split, merge, and rename; pronunciation guides with human voice clips; author and narrator roles; duet seats; live invites; progress tracking; and a named ACX export pack with credits and a retail sample.

Kosmos includes an audiobook-focused recorder, but it is not a general-purpose multitrack music workstation. Narrators can record and punch directly in Kosmos or continue using Reaper, Audition, or another editor and import the result.

Release packaging includes the audited FFmpeg/ffprobe and whisper.cpp runtimes, and Microsoft's offline MarkItDown helper with a built-in fallback. During setup, Kosmos downloads its local speech models once from revision-pinned HTTPS sources, verifies SHA-256 checksums, and keeps them in the user's app-data folder across updates. See third-party notices for model/runtime licenses and attributions.

Install

1. Choose the Mac or Windows download from the Kosmos download page above.
2. On macOS, choose the download for your processor: Apple silicon (M-series) or Intel. The Intel build requires macOS Sequoia 15 or later and uses CPU Whisper for speech checking. Open the download and drag Kosmos to Applications. Release builds are Developer ID signed and notarized by Apple.
3. On Windows, run the installer. If SmartScreen appears, choose More info → Run anyway.
4. On first launch, allow microphone access, choose where project folders live, and download the local proofreading tool. This one-time download is verified before use and is retained through later app updates.

Download the current release once. After setup, recording, proofreading, mastering, and export run locally. Later versions arrive through the in-app updater; restart when you are not recording. Your book folder is unchanged.

Start a book

1. Create a project and choose a manuscript.
2. Choose the local folder where Kosmos will keep the project.
3. Let Kosmos analyze and split the manuscript into chapters.
4. Open a chapter and Record, or import an existing recording.
5. Continue to Proofread, resolve the flags, then open Sound to master and check the chapter.
6. When every chapter is mastered, export the finished book from Export ACX.

ACX mastering and automatic noise cleanup

Export is one click for work Kosmos can perform safely. It checks the selected delivery target, reconstructs detected clicks and clipped peaks, applies conservative steady-noise reduction when needed, normalizes level, limits peaks, resamples and mixes channels, adds compliant room tone, encodes the delivery file, then measures the encoded file again. The app shows one checklist of what was changed and what the delivered file actually passed.

The graphs below come from the deterministic noisy-narration fixture used by verify:delivery. Time runs left to right, frequency runs bottom to top, and brighter colors mean more audio energy. In the first graph, the blue haze across the full height is broadband room noise:

After automatic cleanup, that haze is darker while the bright speech shapes remain in the same places:

FFmpeg independently measures the quiet window moving from −58.8 to −68.8 dBFS while the narration remains −31.4 dBFS before and after. That gives the delivered file more room under ACX's −60 dBFS noise-floor limit without turning down or sanding away the voice. Cleanup is capped at 12 dB; if a take needs more than that, Kosmos refuses to ship it instead of applying destructive processing.

Kosmos uses the same scan → diagnose → repair candidate → preservation check → commit-or-pickup pattern across automatic restoration. Click and clipping repair now runs before noise reduction and mastering; plosive, dereverb, changing-room, and best-take selection workflows are specified in docs/AUTOMATIC_AUDIO_REPAIR.md. A person still confirms the listening experience, but the app prepares the exact pickup only when a bounded repair cannot preserve the human voice.

Offline by design

Manuscript prep, recording, voice follow, proofreading, ACX checking, mastering, and export are local. The app has no sign-in, telemetry, crash phone-home, cloud LLM, voice clone, or remote narrator marketplace. Collaboration is explicit: invite another person to the project or hand off the project folder.

Build from source (contributors)

Requirements: Node.js 20+ and npm. Release packaging also needs the platform toolchain used by Electron Builder.

npm install
npm test
npm run typecheck
npm run dev

Verifying against other tools

npm test checks our code against our own expectations, which cannot catch a mistake we made twice. npm run verify checks the parts that make a claim about the outside world against tools that were not written here, on real audio:

Command	What it proves, and who says so
verify:acx	The mastered MP3 meets ACX's numbers, measured by ffmpeg's astats and volumedetect rather than by us — and an unfixable take is refused instead of shipped.
verify:delivery	FFmpeg's adaptive cleanup lowers steady noise without changing narration level, the cleanup cap is enforced, and an EBU R128 delivery is really 48 kHz mono 24-bit PCM WAV.
verify:restoration	FFmpeg reconstructs planted clicks and clipped peaks, while a clean narration control stays sample- and level-neutral.
verify:loudness	Our LUFS meter agrees with ffmpeg's ebur128 on sines, noise and gated speech at both sample rates.
verify:markers	Every marker export parses as the editor that imports it expects, read back with Python's csv module.
verify:packet	The pickup packet's clips are playable MP3s (ffprobe), its spreadsheet opens in openpyxl, and its page is well-formed HTML.
verify:collab	A pack written by the app's own ZIP writer opens in Python's zipfile and imports back with the right merge, checked against the bytes on disk.
verify:proof	The proof pass scores precision and recall against takes spoken by the macOS voices and decoded by the bundled whisper build.
verify:book	The occurrence scan, the whole-book pickup list and the word filter hold up on those real recordings, and the bundled dictionary's respellings are the ones a narrator wants.

verify:proof and verify:book need the speech model (npm run prepare:model) and macOS say; the rest run anywhere the vendored ffmpeg does. verify:packet also needs Python with openpyxl.

Looking at the panels

Reviewing an interface by reading JSX is guesswork. The design workbench renders the working panels with realistic content — a book mid-proof, a name read two ways, a pack that disagrees with yours — so a layout can be looked at:

npm run design                       # serves design/preview.html
npm run design:shots                 # writes design/shots/*.png
npm run design:shots -- --width 760  # the narrow column, for wrapping

Open http://127.0.0.1:5173/design/preview.html for every panel, add ?panel=pickups for one, and &open to photograph menus open. The page is dev only and is not part of the production entry, so nothing there ships.

For contributors testing a locally built whisper.cpp engine, point the desktop shell at it without changing the project format:

WHISPER_CLI_PATH=/path/to/whisper-cli \
WHISPER_SERVER_PATH=/path/to/whisper-server \
WHISPER_MODEL_PATH=/path/to/ggml-small.en.bin \
npm run dev

For a production bundle:

npm run build
npm run package:mac   # on macOS
npm run package:win   # on Windows

The packaging commands fetch and verify the pinned official Whisper model as a build asset before creating the installer. Release CI also bundles MarkItDown and its DOCX/PDF conversion dependencies; end users do not run either setup step. It records native-runtime checksums and attaches the matching macOS FFmpeg source archive to each GitHub Release.

License

Kosmos is released under the MIT License.