Talk. Get clean markdown. 100% local. 🔗 voice-to-md.dev
Speech-to-Markdown turns your voice into structured documents — and it ships as two apps:
🖥️ Desktop — macOS | 📱 Mobile — iOS | | |---|---|---| | Platform | macOS 13+ (menu-bar app) | iOS 26+, iPhone 15 Pro or newer (Apple Intelligence required; Apple-Silicon iPads too) | | Speech-to-text | |
whisper-cpp
, ffmpeg
, a local LLM servernone— zero setup** 100% offline**— nothing leaves your phoneJump to: 🖥️ Desktop app · 📱 Mobile app
A menu-bar app: voice → whisper.cpp → your local LLM → clean markdown. No cloud. No API keys. Nothing leaves your Mac. Everything in this part — including the dependencies and LLM-server setup below — applies to the Mac app only; the mobile app needs none of it.
↑ This entire product spec was dictated by voice — a local LLM structured it in real time.
⌨️ Global Dictation — Speak, and the transcript is typed straight into whatever field has focus. Terminal, browser, Slack — anything. A Spotlight-style pill shows what's happening:⌘⌥]
anywhere. - 📝 Agent Mode — live markdown editor. Speak freely; a local LLM streams your words into a clean, structured markdown document in real time. Raw transcript stays one click away. Start it from the menu bar:
Everything in Agent Mode is driven from one floating, draggable capsule that hovers over the editor:
Left to right:
— don't wait for the auto-flush: transcribe whatever you just said and send it to the LLM✈️ Send (⌘↩
)right now. Enabled while recording and idle (no LLM call in flight).🎙️ Mic— start, , and resume the session. Red pulse = listening.** Status**—Recording
/Processing
/d
at a glance, with errors surfaced inline.Mode switcher andformat dropdown— how your voice is applied (see below) and what comes out:** MD**(default),** TXT**, or** HTML**. Each format ships its own expectations + example to the LLM, works in every mode, and the session file extension follows (.md
/.txt
/.html
— switching mid-session renames the file). Your choice is remembered across launches.👁️ Preview— opens the session document in the default app for its type (browser for HTML, your editor for markdown/plain text).🗑️ Trash— clears the session (audio, transcript, and document) after a confirmation.
The modes:
| Mode | Icon | What it does |
|---|---|---|
| Format (default) | ||
| 📄 | The whole document + your new words go to the LLM, which returns the complete restructured document. Best for free-form dictation where the model keeps improving the overall structure. | |
| Edit | ||
| ✏️ | Your voice is an instruction, not content: "change the subtitle to Weekly Notes", "turn that list into a table". Select text in the editor first and the model treats it as the focus of the edit. | |
| Append | ||
| ➕ | Speed mode for long documents: only the last 3 sentences + your new words are sent, and the model returns just the new content, which is appended. The LLM never re-reads the whole file, so latency stays flat as the document grows. |
The flow underneath: audio is transcribed by whisper.cpp in ~4 s chunks and buffered; once ~30 words accumulate (or you for 5 s, or hit Send), the buffer is flushed to the LLM using the prompt for the current mode and output format — and tokens stream straight into the editor.
One line — installs dependencies, builds from source, and drops the app in /Applications:
curl -fsSL https://raw.githubusercontent.com/xajik/voice-to-md/main/install.sh | bash
Or from a checkout:
git clone https://github.com/xajik/voice-to-md.git && cd voice-to-md
make install # deps + build + copy to /Applications
First launch: grant Microphone + Accessibility access, then download a Whisper model from Settings… in the menu bar (Base is a great start).
Signing: builds are automatically signed with your "Apple Development" identity when one is in the keychain, so re-installs keep their Microphone/Accessibility grants. No identity → ad-hoc fallback (macOS will re-ask for permissions after updates). Override with make install SIGN_IDENTITY="…"
.
Agent Mode talks to any OpenAI-compatible server. Point Speech-to-Markdown at it in Settings… (default: http://127.0.0.1:8000/v1
, model auto-picked). The Whisper STT model is picked there too:
Pick your server:
Serve any MLX model on port 8000 — Speech-to-Markdown's default, zero config needed:
brew install omlx
omlx serve Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-4bit
brew install ollama
ollama pull qwen3.5 # or: ollama pull gemma3
ollama serve
Base URL: http://127.0.0.1:11434/v1
Download from lmstudio.ai, grab a model, start the local server.
Base URL: http://127.0.0.1:1234/v1
brew install llama.cpp
llama-server -m your-model.gguf --port 8080
Base URL: http://127.0.0.1:8080/v1
| Model | Why |
|---|---|
| Qwen3.5 27B (4-bit) | |
| Best formatting quality; the default pick | |
| Gemma 26B (8-bit) | |
| Fast, great instruction following |
Anything that follows instructions well works — Speech-to-Markdown streams tokens as they arrive, so even bigger models feel instant.
Agent Mode on your phone — everything runs offline, on the device. No LLM server, no whisper install, no settings, no network. None of the desktop dependencies above apply here.
Requirements:iOS 26 or lateron anApple Intelligence device — iPhone 15 Pro or newer(or an Apple-Silicon iPad). Apple Intelligence must be enabled in Settings; the app shows an actionable banner if it isn't.
STT: Apple's SpeechAnalyzer streams your speech to text live, fully on-device (volatile text shows in gray as you talk; finalized words feed the document).LLM: Apple** Foundation Models**— the on-device Apple Intelligence model — formats the transcript. No server, no API keys, works in airplane mode.- Same three modes ( Format / Edit / Append), same** MD / TXT / HTML**output formats, sessions restorable from history, documents visible in the Files app.
Building from source: make build-ios
/ make test-ios
(see Development). Running on a device needs a development team: xcodebuild -scheme SpeechToMarkdownIOS -allowProvisioningUpdates CODE_SIGN_STYLE=Automatic DEVELOPMENT_TEAM=<TEAMID> …
(and Developer Mode enabled on the phone).
make check # verify dependencies
make build # release build + sign
make test # unit tests
make generate # regen .xcodeproj after editing project.yml
make install # build + install to /Applications
make uninstall # remove from /Applications
make build-ios # build for the iOS simulator
make test-ios # run the iOS test suite
Both apps share one core (Shared/
): transcript buffering, prompts, formats, session files. Desktop pipeline: AVAudioEngine → whisper.cpp → local LLM server. Mobile pipeline: AVAudioEngine → SpeechAnalyzer → Foundation Models. Either way, everything stays on your hardware.