What it actually does
Six stages, each one resumable and written to plain files on disk, so you can stop after any step and pick the project back up later.
Step 1
Source material
Paste a YouTube link and the app pulls the subtitles as reference, or paste your own text. Optional — you can start from a blank idea.
Step 2–3
Brief & script
The app builds a prompt pack you paste into ChatGPT. Paste the answer back and it extracts the script, the self-scored review and the metadata.
Step 4
Originality check
The engine measures the script against your sources itself — word-overlap and meaning-similarity — instead of taking the AI's word for it.
Step 5
Narration
Two Vietnamese voices ship with the app, or clone your own from a few seconds of audio. The model runs locally on your CPU.
Step 6
Export
MP3 and WAV, a matching SRT subtitle file, and a description file with the source credits already written in.
Optional
Claude Desktop
If you use Claude Desktop, the app can expose the project to it over a local MCP server so the assistant writes straight into it — no copy-paste.
Before you install
The installers are not code-signed yet
Kể is free and self-funded, and a code-signing certificate is not in place. Both Windows and macOS will therefore warn you the first time you open it. That warning means “this developer has not paid for a certificate”, not “this file is dangerous” — but you should only trust that from a source you trust, so the SHA-256 of each file is published above and you can verify it before opening.
Opening it on Windows
- Run
Ke-Setup-0.2.2-win-x64.exe. - SmartScreen will say “Windows protected your PC”. Click More info, then Run anyway.
- Install. It does not need administrator rights and installs for your user only.
Verify the download first, if you like — in PowerShell:
Get-FileHash .\Ke-Setup-0.2.2-win-x64.exe -Algorithm SHA256
Opening it on a Mac
- Open
Ke-0.2.2-arm64.dmgand drag Kể into Applications. - The first launch will be blocked. Open System Settings → Privacy & Security, scroll down, and click Open Anyway next to the message about Kể.
- If that option does not appear, run this in Terminal and open the app again:
xattr -dr com.apple.quarantine /Applications/Kể.app
To verify the download:shasum -a 256 Ke-0.2.2-arm64.dmg
Apple Silicon only
The macOS build runs on M1, M2, M3 and M4 Macs. There is no Intel build — on an Intel Mac it will not start. Check with the Apple menu → About This Mac; it should say “Apple M…”, not “Intel”.
First launch downloads about 1.2 GB
The speech engine is not inside the installer. The first time you use step 5, the app installs a small Python runtime (about 620 MB) and the voice model (about 610 MB), once, over your internet connection. Everything after that works offline.
What you need
| Windows | Windows 10 or 11, 64-bit. About 1.5 GB free disk space including the voice model. |
|---|---|
| macOS | Apple Silicon (M1–M4), macOS 13 or newer. About 1.5 GB free disk space. |
| Internet | Needed to pull YouTube transcripts and for the one-time voice model download. The rest works offline. |
| An AI subscription | Your own ChatGPT (or Claude) account for the writing step. Kể has no AI of its own and never asks for an API key. |
| Rendering time | Narration runs at roughly 1.2–1.7× realtime on an M4, so a 30-minute episode takes 18–25 minutes. It runs as a background job with progress. |
Your text and your voice stay on your computer
Kể has no backend. There is no account system, no telemetry, and no server for it to send anything to. Projects live in a folder you choose; cloned voices live in your own application data folder. The only network traffic is YouTube subtitle downloads, the one-time model download, and — if you use it — whatever you personally paste into ChatGPT.
How it was built
We wrote up the engineering behind this one — the originality gate and how its threshold was calibrated, why there is no API key in the app, what local speech synthesis actually costs, and the Windows-only transcript bug that three of our tests were quietly mocking out.
Read the case study →