DCENT_Voice v0.2.0-beta.1 — Source Released
DCENT_Voice: Open-Source, Local-First Dictation
Hold a hotkey, speak, and release. DCENT_Voice transcribes on your machine, polishes the words locally, and types them into the app you already have open. The public-beta source is live on GitHub now — no account, no subscription, and no silent cloud fallback.
The Source Is Live
DCENT_Voice is now public on GitHub and already serves as a daily driver at D-Central. It is a real, usable beta with a substantial local-first feature set — and it still carries a beta label because rough edges and changing defaults are honest possibilities.
Public Beta
v0.2.0-beta.1 is the first public version. Windows and Linux are supported today; macOS is implemented and CI-tested but remains a preview until it receives the same hardware soak testing.
MIT-Licensed Source
The complete source is public under the MIT License. Read it, audit it, build it, fork it, or help improve it. Bugs and feature requests belong in the public issue tracker.
Start on GitHub
The repository is the canonical install source. Its README carries the current Windows, Linux, macOS-preview, source-install, diagnostics, and model instructions in one place.
Talk Where You Would Normally Type
The default workflow is deliberately simple: put the cursor in an email, chat, editor, terminal, note, or web form; hold Ctrl+Win; speak; release. DCENT_Voice transcribes the utterance, cleans it, inserts it into the focused app, restores your previous clipboard, and returns to idle.
Push to Talk
Hold the global hotkey while speaking and release to finish. A voice-reactive overlay shows listening, transcription, command mode, and successful insertion without stealing focus.
Command Mode
Use spoken intent around selected text and editing tasks. DCENT_Voice prefers native selection access when available and falls back safely when an app exposes less control.
Streaming Dictation
For longer thoughts, streaming mode types stable words while you speak. Finalization reconciles the tail so an in-flight transcription does not swallow the end of the utterance.
What Ships in the Public Beta
- On-device speech recognition by default — CPU ONNX Parakeet TDT 0.6B v3 is the desktop default, with bundled Faster Whisper as the offline fallback.
- Local writing polish without an LLM — filler removal, false starts, mid-utterance corrections such as “five, actually six,” spoken structure cues, snippets, and destination-aware style run locally by default.
- Optional local-AI cleanup — Ollama or LM Studio can perform an additional cleanup pass when you enable it; raw transcription still works when that service is unavailable.
- System tray, setup wizard, and diagnostics — pick and test a microphone, see the active hotkey, inspect model health, and run the same self-check used by Windows Setup.
- Safe text injection — per-app injection overrides accommodate terminals and remote apps, while clipboard-paste injection preserves text, rich text, images, and copied-file formats when they can be cloned safely.
- Recovery when insertion fails — an optional, bounded recovery vault can retain usable text that failed to insert. It is off by default and never stores successful dictation or microphone audio.
- Local API and DCENT_ADE attachment — a token-secured localhost service exposes transcription, commands, events, streaming, and the DVAP attachment protocol without opening a general cross-origin endpoint.
- Optional cloud providers with an explicit gate — OpenAI, Groq, xAI, Anthropic, and Deepgram are available only after provider selection and recorded consent. Use is visible in the privacy state and metadata-only egress log; cloud is never a silent fallback.
A GPU Is Optional
The shipped desktop profile is a CPU path, not a consolation fallback. You can change models when your language, hardware, or accuracy target calls for it.
| Profile | Local engine | Best fit |
|---|---|---|
| desktop (default) | CPU ONNX Parakeet TDT 0.6B v3 | Fast automatic decoding across Parakeet’s documented 25 languages |
| multilingual | Faster Whisper base, CPU int8 | Corpus comparisons and languages outside Parakeet’s documented set |
| tiny / quality | Faster Whisper tiny.en / distil-small.en | Older laptops or better English accuracy at a slower finalize |
| auto / gpu / accurate | CUDA when ready, otherwise local CPU | Known-good NVIDIA systems or a deliberate quality-first profile |
Local text-to-speech with Kokoro is a source-preview feature. Its optional runtime is not bundled in the Windows public beta while license compatibility is reviewed, and it advertises no TTS capability until a compatible runtime and verified assets are actually present.
Platform Status
| Platform | Status | What to know |
|---|---|---|
| Windows 10 1809+ / 11, x64 | Supported | Most-exercised path. The native Setup design is per-user: no admin prompt, Python installation, or separate model download. |
| Linux, X11 and Wayland | Supported | Capture, tray, overlay, and injection are implemented; source installation requires the documented desktop and injection packages. |
| macOS | Preview | Code paths and CI coverage exist, but real-hardware soak testing has not yet matched Windows and Linux. |
Native public-beta artifacts are unsigned. Follow the repository’s current install instructions and verify any published SHA-256 before running a download.
Local First Means You Can See the Boundary
The normal dictation path keeps microphone audio and transcription on your hardware. No account makes it better, and losing the internet does not break local speech recognition.
Local Is the Default
Transcription, deterministic writing cleanup, settings, snippets, and the local service run on the machine. Optional LLM cleanup is off until you enable it.
Cloud Requires Consent
A cloud provider cannot start until it is selected and its consent record exists. Provider keys live in the operating-system keychain, not in the TOML configuration.
Egress Is Visible
The interface reports sovereign, hybrid, or cloud posture. The egress log records provider, payload type, timestamp, and byte count — never the dictated content itself.
Frequently Asked Questions
Is DCENT_Voice open source now?
Yes. The DCENT_Voice repository is public under the MIT License. The current project version is v0.2.0-beta.1.
Does it need a GPU?
No. The default desktop profile runs Parakeet through CPU ONNX. Faster Whisper also has CPU profiles, and CUDA acceleration is optional for a machine with a known-good NVIDIA runtime.
Does my voice ever leave the computer?
Not on the default local path. Optional cloud integrations exist, but they require an explicit provider choice and consent record, display the changed privacy posture, and never act as a silent fallback.
What languages does it support?
The default Parakeet TDT 0.6B v3 profile documents 25 European languages and automatic detection. The bundled Faster Whisper fallback covers additional language needs through its multilingual profiles.
Can I use it with DCENT_ADE?
Yes. DCENT_Voice exposes a token-secured local API and DVAP attachment envelope for transcription events and ADE integration. Optional source-preview TTS capabilities appear only when a compatible local runtime and verified assets are installed.
Where should I report a rough edge?
Open an issue in the GitHub issue tracker with your platform, the relevant diagnostics check, and the smallest reproducible description. Please do not include API keys, dictated content, or private logs.