Topics

#dictation #llm #local-first #macos #metal #open-source #privacy #productivity #rust #speech-to-text #tauri #voice-dictation #whisper

Language breakdown

No language data available.

▸ Files click to expand

README

Sumi

System-wide speech-to-text for macOS. Free and open source.

Press a hotkey anywhere. Speak. The text pastes at your cursor — rewritten by an AI that knows which app you're in.

English | 繁體中文 | 简体中文


_{Gmail — formats as a proper email}	_{Notion — clean structured prose}	_{Telegram — casual, conversational}

brew tap alan890104/sumi && brew install --cask sumi

Download DMG · Releases · Issues

Why Sumi?

🎯 Per-App AI Polish

Every app gets a different LLM prompt. Slack sounds like Slack. Gmail sounds like email. Terminal gets clean commands. 18 built-in rules — write your own, or describe what you want and the AI generates the rule.

🔒 Fully Local

Run everything on-device: Whisper or Qwen3-ASR for speech recognition, local LLM for rewriting. Audio never leaves your Mac. You can verify it — the code is here.

🗣 Speaker Diarization

Meeting mode transcribes continuously in the background. Transcripts are labelled by speaker with timestamps. Import existing audio files for retroactive transcription.

✏️ Edit by Voice

Select any text, press Option+E, say what you want done. "Make this more formal." "Translate to Japanese." "Shorten it." The AI rewrites and pastes back.

☁️ Cloud or Local — Your Choice

BYOK for everything: Groq, OpenAI, Deepgram, Azure, Gemini, OpenRouter, or any OpenAI-compatible endpoint. No Sumi account. No subscription.

🌏 58 UI Languages

The interface ships in 58 locales. Traditional Chinese users get automatic zh-CN → zh-TW normalization on transcription output.

Same sentence, three apps

You say: "um I think the project is kind of behind schedule and we should probably have a meeting to figure out what to do next"

LINE (casual)	I think the project is behind schedule, we should have a meeting to figure out what to do next
Slack (professional)	I think the project is behind schedule. We should have a meeting to discuss next steps.
Gmail (email)	Hi, I believe the project is currently behind schedule. Could we schedule a meeting to discuss the next steps? Best regards

How it works

App lives in the menu bar — nothing else on screen.
Click into any text field, anywhere on your Mac.
Press Option+V. A floating capsule appears with the waveform.
Speak.
Press Option+V again. Text pastes.

Edit by Voice: select text → Option+E → say what to do with it. Meeting Mode: press Option+M to toggle continuous background transcription into a note file.

What it runs on

Speech recognition — Local: Whisper (Metal GPU, 7 model sizes from 148 MB to 1.6 GB) or Qwen3-ASR. Cloud: Groq, OpenAI, Deepgram, Azure, any custom endpoint.

LLM rewriting — Local: Qwen3-8B, Qwen2.5-7B, Llama 3 Taiwan 8B via candle (Metal/CUDA). Cloud: OpenAI, Groq, Gemini, GitHub Models, OpenRouter, SambaNova, any OpenAI-compatible endpoint.

Resource usage — Idle: ~130 MB, 0% CPU. Local transcription: ~730 MB RSS, <20% CPU (Metal). Cloud mode: ~7 MB during recording, back to 0% when done.

Other details — Silero VAD for silence filtering · custom pronunciation dictionary · transcription history with audio export · customizable hotkeys

Comparison

[!NOTE]

This reflects our best understanding at the time of writing. Competitors update frequently — corrections welcome via issues or PRs.

| | Sumi | Built-in Dictation | Typeless | Wispr Flow | VoiceInk | SuperWhisper | |---|---|---|---|---|---|---| | Price | Free | Free | 4K words/wk free, $12-30/mo | 2K words/wk free, $12-15/mo | $25-49 (one-time) | Free trial, ~$8/mo | | Open Source | ✅ GPLv3 | ❌ | ❌ | ❌ | ✅ GPLv3 | ❌ | | Local STT | ✅ | ✅ Apple Silicon | ❌ Cloud only | ❌ Cloud only | ✅ | ✅ | | Cloud STT | ✅ BYOK | ❌ | ✅ | ✅ | ✅ Optional | ✅ | | AI Polish | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | | Local LLM Polish | ✅ 3 models | ❌ | ❌ | ❌ | ❌ | ✅ | | Per-App Rules | ✅ 18 presets + custom | ❌ | ❌ | ✅ Styles | ✅ Power Modes | ✅ Custom modes | | Context-Aware | ✅ App + URL | ❌ | ✅ App | ✅ App | ✅ App | ✅ Super Mode | | Edit by Voice | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | | Dictionary | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | | History | ✅ + audio export | ❌ | ✅ | ✅ | ✅ | ✅ | | Meeting notes | ✅ + speaker labels | ❌ | — | ❌ | — | ✅ | | Platforms | macOS | macOS, iOS | macOS, Win, iOS, Android | macOS, Win, iOS, Android | macOS | macOS, Win, iOS |

Installation

Homebrew (recommended)

brew tap alan890104/sumi
brew install --cask sumi

Download DMG

Download the latest DMG from GitHub Releases.
Open the DMG, drag Sumi into /Applications.
The app isn't notarized yet, so macOS will block it. Run this first:

xattr -cr /Applications/Sumi.app

On first launch: grant Microphone access and enable Accessibility under System Settings → Privacy & Security → Accessibility (required for auto-paste).

Build from source

Requires Rust and cargo install tauri-cli --version "^2".

git clone https://github.com/alan890104/sumi.git
cd sumi
cargo tauri dev      # dev mode
cargo tauri build    # production .dmg

Windows (CUDA)

Metal is macOS-only. On Windows:

# CPU only
cargo tauri dev --no-default-features
NVIDIA CUDA (requires CUDA Toolkit, LLVM, Ninja, CMake)
bash dev-cuda.sh
bash dev-cuda.sh --release

License

GPLv3

🔗 More in this category