Dictovicto is out: transcription that never phones home

#announcement

Dictovicto is out: transcription that never phones home

Today we're releasing Dictovicto, a transcription and dictation studio for Windows that runs entirely on your PC. Not "mostly local with a cloud fallback." Not "local, but sign in first." Entirely on your PC. You could pull the network cable out of the wall mid-transcription and Dictovicto would not notice.

It's available now on the Microsoft Store.

Why we built it

Every transcription tool we tried had the same fine print: your audio gets uploaded somewhere. For a podcast episode, maybe that's fine. For a client interview, a medical dictation, a legal deposition, or your company's quarterly planning meeting? Some audio should never touch someone else's server, full stop.

The usual answer is "run Whisper yourself," which in practice means a weekend of Python environments, CUDA version roulette, and a command line where an app should be. We wanted the privacy of local inference with the polish of a real product. So we built one.

What it does

Live dictation. Talk, and watch the words appear as you speak.
Import audio or video. MP3, WAV, M4A, FLAC, OGG, MP4, MKV, WebM, and more — drop the file in and Dictovicto handles the decoding itself, no ffmpeg install required.
Speaker diarization. See who spoke when. Rename speakers, and fix any misattributed lines with a click.
A real transcript editor. Click a timestamp to replay that moment, edit text inline, delete lines. Built for the cleanup pass, because every transcript needs one.
Export anywhere. Plain text, SRT and WebVTT subtitles, CSV, and structured JSON.
Five Whisper models bundled, from the lightning-fast tiny to the state-of-the-art large-v3.
Hardware-aware. Dictovicto probes your machine on first launch and recommends the best model it can actually run — on an NVIDIA GPU if you have one, on your CPU if you don't.

The part we're stubborn about

Every AI model ships inside the app. All five Whisper tiers, the diarization models, the GPU runtime — everything is in the installer. Nothing downloads after install, ever.

Yes, that makes the package big. Multi-gigabyte big. We made that trade on purpose: a large install that's guaranteed to work offline forever beats a slim one that quietly needs a CDN to become useful. Install it once, transcribe until the heat death of the universe.

And to be precise about "offline," since everyone claims it: the only network call the app ever makes is to the Microsoft Store's licensing API to confirm your purchase — and even that is usually answered from Windows' local license cache. No account. No subscription. No telemetry. The only thing that ever leaves Dictovicto is the text you choose to export.

What you'll need

Minimum: Windows 10 (64-bit, version 1809 or later), a quad-core x64 processor, 8 GB of RAM, 16 GB of free storage, and a microphone if you want dictation.

Recommended: an 8-core CPU, or an NVIDIA GPU with 6 GB+ of VRAM if you want the large-v3 model at full speed. GPU acceleration is currently NVIDIA-only (CUDA); AMD and Intel GPUs will fall back to the CPU, which works fine — the smaller models are genuinely quick even without a GPU.

Under the hood, briefly

For the curious: Dictovicto is a Tauri v2 app with a Svelte frontend and a native Rust inference engine built on sherpa-onnx and ONNX Runtime. The entire ML stack lives in an isolated sidecar process that talks to the app over a line-delimited JSON protocol — the same Whisper and diarization models the research world uses, minus the multi-gigabyte Python runtime they usually drag along. Audio decoding is pure Rust via symphonia, and live capture goes straight through WASAPI without ever touching the browser layer.

We'll write more about the engine in a future post, including the saga of replacing a Python sidecar with an 18 MB Rust executable. There was swearing. It was worth it.