DCENT_Voice — D-Central Technologies

DCENT_Voice: Local, Private Voice-to-Text

Press a key, talk, and clean text appears wherever your cursor is — email, chat, code, notes, anywhere you type. Open-source dictation that runs entirely on your own machine with local models. No cloud, no telemetry, no per-seat subscription. In public beta; the source lands on GitHub once polishing is done.

Follow the Beta See the DCENT Suite

100%

On-Device Transcription

Cloud Calls

99+

Languages (Whisper)

Per-Seat SaaS Fee

Public Beta — Source Lands Once Polishing Is Done

DCENT_Voice is in public beta, tested daily on real hardware while we polish it. Here is the honest status — we would rather under-promise.

Where It Is

In public beta. We are dogfooding it daily and shaking out the last rough edges before the source goes public.

Source Release

Like the rest of the DCENT suite, DCENT_Voice will be open source under GPL-3.0. The repository joins github.com/DCentralTech once polishing is done — read it, audit it, build it yourself.

Public Beta

The public beta is open now — a test, not a finished product. Join the Discord to take part and follow the build.

What Is DCENT_Voice?

Cloud dictation apps like Wispr Flow proved how good this workflow feels: hold a hotkey, talk, and polished text lands at your cursor in any app. They did the hard work of proving the pattern — and we are grateful for it. DCENT_Voice takes that same experience and moves it one more layer decentralized: the microphone, the transcription, and even the clean-up all run on hardware you already own. Your voice never becomes someone else’s data.

The Same Great Workflow

System-wide, push-to-talk dictation. Speak naturally; DCENT_Voice transcribes, punctuates, removes the “ums”, and inserts clean text wherever you were typing — Gmail, Slack, your editor, a terminal, a web form.

The Difference: Nothing Leaves

Cloud dictation streams your audio to someone else’s servers, needs an internet connection, and meters you per seat. DCENT_Voice runs the whole pipeline locally: no upload, no account, no subscription, works on a plane.

Built for Plebs

The same sovereignty logic behind Bitcoin and self-hosted AI, applied to your keyboard. Own your money, own your compute, own your words. It fits naturally into the sovereign AI stack we already document.

Standing on the Shoulders of Giants

DCENT_Voice invents almost none of the hard science. It is a careful integration of the open-source speech and language work that made local dictation possible. Credit where it is due.

The Speech Engine

OpenAI’s Whisper (MIT) is the foundation model. Georgi Gerganov’s whisper.cpp runs it on almost anything, and SYSTRAN’s faster-whisper flies on a GPU. Accuracy comes from their work; we just wire it in.

The Clean-Up Layer

Punctuation and formatting are polished by a local model via Ollama / llama.cpp (Gerganov again) — so even the “make it read well” step stays on your machine. Local voice activity and wake detection lean on Silero VAD and openWakeWord.

Fellow Travellers

We are not first, and we will not pretend to be. Projects like Handy, OpenWhispr, and Superwhisper already do local dictation well. DCENT_Voice’s angle is tight integration with the wider sovereignty stack — more good options, not fewer.

What It Does

The feature set we are building toward. Some of this is live in the beta today; some is still landing.

Fully on-device transcription — audio never leaves the machine; no cloud, no telemetry.
System-wide hotkey dictation — clean text at the cursor in any app.
Local LLM post-processing — punctuation, capitalization, filler-word removal, and formatting handled by a local model, so the polish step is private too.
Model choice, tiny to large-v3 — trade accuracy against speed to match your hardware, from a Raspberry Pi to a GPU workstation.
Offline by default — after a one-time model download, no internet required.
No per-seat fee, no lock-in — GPL-3.0, yours to run forever.
Multi-language — Whisper covers 99+ languages out of the box.
Optional local wake word + VAD — hands-free trigger without an always-listening cloud mic.

Honest Requirements & Trade-Offs

Local means the work happens on your box, so your hardware sets the ceiling. No magic — just physics you control.

Model tier	Runs comfortably on	Trade-off
tiny / base	Modern CPU, Raspberry Pi 5, any laptop	Fastest; lower accuracy on hard audio
small / medium	A recent CPU or entry GPU / Apple Silicon	Good balance for everyday dictation
large-v3 / distil-large	GPU with ~6–10 GB VRAM or Apple Silicon	Best accuracy; heavier and slower on CPU

Accuracy is inherited. DCENT_Voice runs the same Whisper weights as every other local tool — transcription quality tracks the model you pick, not our branding.
Latency is hardware-dependent. On a decent GPU or Apple Silicon box it feels instant; a large model on a slow CPU will not.
It is beta. Expect rough edges, changing defaults, and the occasional bug while we test.

Frequently Asked Questions

Is it really fully local?

Yes. Transcription runs on-device with the Whisper family, and even the clean-up pass uses a local model. There is no cloud endpoint in the pipeline and no telemetry. After you download a model once, it works offline.

How is it different from Wispr Flow or Otter?

Those are excellent cloud products — they proved the workflow. The core difference is architectural, not a knock: they transcribe in the cloud on a subscription; DCENT_Voice transcribes on your own machine, open-source and free to run. Different trade-offs; pick what fits your threat model.

What hardware do I need?

Anything from a Raspberry Pi 5 up. Small models run on a plain CPU; the largest, most accurate models want a GPU with roughly 6–10 GB of VRAM or Apple Silicon. You choose the balance of speed and accuracy.

Is it free and open source?

Yes — GPL-3.0, like the rest of the DCENT suite. There is no per-seat fee and nothing to unlock. The source lands on GitHub once polishing is done; the beta is public in the meantime.

Can I try it now?

Yes — the public beta is open while we polish. Join the Discord to get set up, and you will be first to the source drop on GitHub.

Who builds DCENT_Voice?

D-Central Technologies — Bitcoin mining hackers who like owning their tools. We build on Whisper, whisper.cpp, faster-whisper, Ollama, and the local-AI community; our job is the integration and the polish, not the underlying science.

Related products, repair, and setup paths

Last reviewed July 22, 2026.