Show HN: Dictata – Local Whisper dictation with LLM cleanup

100% local system-wide voice dictation for Windows, in native Rust. Press a global hotkey, speak, press again: the text is transcribed locally by whisper.cpp and pasted into the active application. No data ever leaves the machine.

Current release: v0.1.0 — see CHANGELOG.md.

Hotkey dictation(toggle or push-to-talk), automatic paste into the active application, floating dock with waveform (configurable size, opacity and position).Continuous mode (streaming): text is inserted as you speak, at every detected .** Silence skipping (VAD)**: optional whisper.cpp voice-activity detection on standard (non-streaming) dictation — less compute and fewer hallucinations on silence. Downloads a small (~2 MB) model on first use; toggle in Configuration.Vulkan GPU transcription(AMD/Intel/NVIDIA) or CPU — whisper-rs with thevulkan

feature.Audio sources: microphone, system audio (WASAPI loopback) or a mix of both (meeting mode).** Output modes**: raw, or post-processing through a local OpenAI-compatible LLM (cleanup, email, message, list…), with customizable prompts.Transcribe a file(audio/video, History page) — any format handled by ffmpeg.- Custom vocabulary and replacements injected as the initial prompt. - Built-in ggml model library: download, delete, and hardware-aware recommendations (detected GPU/VRAM when running on Vulkan, CPU/RAM otherwise). HuggingFace search: install any ggml.bin

model from a direct URL, a repo (owner/name

) or a keyword search.fr / en / es interface, dark theme, system tray icon, transcription history.

• Windows 10/11. Linux is not supported in this release — seeLINUX.mdfor what works, what does not, and why. ffmpeginPATH

(only for "Transcribe a file").- To build with GPU support: the Vulkan SDKand the Visual Studio Build Tools (CMake + Ninja included).

cargo build --release

Build specifics:

MAX_PATH workaround (Windows): whisper.cpp'svulkan-shaders-gen

sub-project can exceed the Windows MAX_PATH limit (260 characters) and break MSBuild (FTK1011/C1083). If that happens, create a local (untracked).cargo/config.toml

pointing the build to a short path at a drive root:

[build]
target-dir = "C:/wt"   # any short path

The executable and its

config.json

then live in<target-dir>\<profile>\

. - After a

cargo clean

, rebuild inside a Visual Studio environment with the Ninja generator:

VsDevCmd.bat -arch=amd64 && set CMAKE_GENERATOR=Ninja && set CMAKE_GENERATOR_INSTANCE= && set VULKAN_SDK=<SDK path> && cargo build

For a CPU-only build (no Vulkan SDK required):

cargo build --release --no-default-features
powershell -ExecutionPolicy Bypass -File scripts\release-windows.ps1 -Variant both

Runs the tests, builds, and writes dist\dictata-<version>-windows-x86_64[-cpu].zip

with its SHA-256. The zip contains only dictata.exe

, README.md

, LICENSE

and CHANGELOG.md

: no config.json

is shipped (the app writes its own defaults on first run, and a development config.json

carries personal data). The binary is self-contained — no DLL to ship alongside it; the Vulkan build uses the provided by the graphics driver.

The executable carries its icon and version metadata, embedded by build.rs

via winresource

. The icon itself is drawn by a script rather than checked in as an opaque blob — edit the numbers and regenerate all sizes at once:

powershell -ExecutionPolicy Bypass -File scripts\make-icon.ps1

scripts/release-linux.sh

mirrors the packaging script, but has never been run — see LINUX.md.

• Launch the executable: the application lives in the system tray.
• Open Settings(tray menu, or set theDICTATA_OPEN_SETTINGS=1

environment variable at launch): pick a model in the Models page (built-in download), set the hotkey, language and audio source. - Place the cursor where you want to write, press the hotkey (default Ctrl+Alt+Space

), speak, press again.Hold for about half a second to cancel the current take — Escape is read globally, so a plain tap is left to the application you are typing into.Esc

The configuration is read from config.json

next to the executable, or from the directory pointed to by the DICTATA_HOME

environment variable. The file must be UTF-8 without BOM.

GPU: gpu

config field — "auto"

(default, uses Vulkan when available), "cpu"

, "vulkan"

, "cuda"

.

Module	Role
main.rs
Orchestration: states, hotkey, dock, transcription threads
audio.rs
cpal capture (mic / loopback / mix), 16 kHz resampling, ffmpeg decoding
transcriber.rs
whisper-rs wrapper (model , transcription)
streaming.rs
Continuous mode: -based chunking, progressive emission
settings.rs
Settings window (8 pages, egui) — presentation
settings_logic.rs
Settings business logic, testable without UI
dock.rs
Floating dock (waveform, states)
modes.rs / llm.rs
Output modes and local LLM post-processing
models.rs
ggml catalog, download/delete, HuggingFace search, paths
i18n.rs
fr/en/es translations (tr() )
config.rs / history.rs / paste.rs / hotkey.rs / tray.rs / hardware.rs / platform.rs
Config, history, paste, global hotkey, tray, hardware detection, Windows integration

cargo test

64 unit tests (config, resampling, mixing, modes, settings logic, hardware rating, HuggingFace query parsing, audio-context sizing, history retention, cancel gesture, download safety…). A CLI example is provided:

cargo run --example transcribe -- <file.wav>   # DICTATA_GPU=1 for GPU

Licensed under the MIT License with the Commons Clause — see

LICENSE.

Free to use, copy, modify, distribute and self-host, for personal or any other purpose.You may not: you cannot sell the software, or offer a paid product or service (hosting, SaaS, paywall, paid support…) whose value derives substantially from it.Sellit

© 2026 Antoine Chatry.

source & further reading

github.com — original article Talos AI Agent Super Secure Show HN: Macsomnia – close your MacBook lid while keeping long jobs alive Show HN: Kvcachescope – Why Nvidia-smi is blind to vLLM KV cache leaks