Hold a key, speak, and the text lands where your cursor is. No account, no server, no telemetry. Audio and text go only where you point them: an on-device model, or an endpoint you configure.
Features
- On-device models. Parakeet v3, Parakeet Unified, Nemotron English, Whisper and Cohere run entirely on your computer, using ggml (transcribe.cpp), with GPU acceleration and Silero VAD built in.
- Bring your own key. Point the app at any OpenAI-compatible speech endpoint (Groq works well). Your key lives in the system keychain.
- AI editing. Optional cleanup removes fillers and false starts and formats lists and emails: on device in English with S1-mini by Superwhisper, or through your own key. With your own key, command mode edits selected text or answers a spoken question, and translation covers any language.
- Works everywhere you type. evdev hotkeys and a layer-shell overlay on Wayland, CGEventTap on macOS, a low-level hook on Windows. Injected text is verified against the field and falls back to the clipboard, so a dictation is never lost.
- Notes, dictionary, snippets. Quick local notes, custom vocabulary that biases recognition, text expansion, per-app styles, and corrections learned from your edits.
Footprint
A native Rust and GPUI app, not Electron. Installers are 11 to 24 MB.
| App | Memory at idle | CPU at idle | Source |
|---|
| TypeAside | 30 MB (85 MB window open) | 0% | Measured |
| TypeAside, Parakeet v3 loaded | 850 MiB, freed after 30 min idle | 0% | Measured |
| TypeAside, Parakeet v3 + S1-mini by Superwhisper | 1,680 MiB PSS, freed after 30 min idle | 0% | Measured |
| Wispr Flow | 750 to 800 MB | about 8% | tl;dv test, user reports |
| Superwhisper, local model loaded | about 900 MB | 0% | User report |
With Parakeet v3 Q4_K_M on the CPU, 9.75 s of speech transcribes in 325 ms using 2.46 CPU-seconds; 29.25 s takes 1,006 ms and 7.72 CPU-seconds. Worker PSS is 850 to 851 MiB loaded and 864 to 891 MiB after transcription, with 0% idle CPU. These are five-run medians from the real worker on Linux, Ryzen 9 8945HS, eight inference threads, transcribe.cpp 0.4.1. The previous four-thread ONNX baseline was 280/950 ms and 1.37 to 1.53 GB PSS: GGUF reduces memory and download size with a modest latency cost in this worker measurement. See model verification for details. Competitor figures are third-party and vary by platform and version.
With Parakeet v3 and S1-mini by Superwhisper loaded together, the CPU worker uses 1,680 MiB PSS after the first cleanup and 1,708 MiB after dictations. For a 4.88 s clip, speech takes 156 ms and cleanup adds 385 ms, 2.70 CPU-seconds in total; a 9.75 s clip takes 314 + 431 ms and 4.07 CPU-seconds. Five-run medians on the same machine, eight speech and four cleanup threads, 0% idle CPU. See offline AI editing for outputs and limits.
Install
| Platform | Command |
|---|
| Linux | curl -fsSL https://typeaside.app/install.sh | sh |
| Arch | yay -S typeaside-bin |
| macOS | brew install --cask adityamiskin/tap/typeaside |
| Windows | Installer from typeaside.app/download |
Screenshots
Development
The desktop app is Rust and GPUI. Building needs CMake and, on Linux x86_64, Vulkan headers.
cargo check
cargo test
cargo run
The website (apps/web) is TanStack Start on Cloudflare Workers, built with Bun and Vite+. It needs no credentials to run.
bun install
bun run dev:web # http://localhost:3001
bun run check # format, lint and type checks
bun run build
| Path | What lives there |
|---|
src/, assets/, vendor/ | Desktop app |
apps/web | Marketing site, docs and downloads |
packages/ui | Shared shadcn primitives and the Tailwind v4 theme |
docs/ | Platform notes, model catalog, history limits |
examples/ | Runnable end-to-end checks for the pipeline and local models |
