Hold the hotkey → speak → finished text lands in the active app.
No cloud service, no subscription, no third-party app. 100 % offline.
Your audio never leaves your Mac. whisper.cpp runs directly on your hardware — no cloud, no server, no outbound calls.
Build once, use forever. No sign-up, no subscription, no tracking. The tool is yours.
Full source code on Codeberg. An independent implementation built on whisper.cpp, FluidAudio (Parakeet, Apache 2.0) and Ollama.
A native macOS menu bar app — lean, precise, no bloat.
Choose your engine: Whisper (best quality), Parakeet (near-instant on the Neural Engine, 25 European languages) or Apple Speech (on-device, no download). Settings → Models explains the trade-offs — all run locally. Downloaded models are checked against a known checksum.
Default: double-tap — doesn't conflict with ⌥ special characters like @. Alternatively push-to-talk (hold) or the fn/🌐 globe key. Freely assignable shortcut, e.g. ⌥ Space. Esc discards the recording.
Raw · Clean up · As email · Translate to English — plus custom actions in the editor. Optionally use a cloud provider (Langdock / OpenAI-compatible) instead of Ollama locally.
"Keep model in memory" starts whisper-server in the background — transcription begins instantly, no model load time on first dictation.
Beta (Parakeet): the recording window shows your words as you speak — waveform, auto-scroll, incoming text. What finally gets inserted is always the full, clean result.
Automatic word replacement (dictionary.txt), a proper noun list (vocabulary.txt) for the LLM, and text snippets via trigger word. New: optional text cleanup without any AI — strip filler words and turn spoken punctuation into symbols.
Recent transcriptions accessible in the window and directly from the menu bar menu. Search, pin, delete individually, drag & drop out.
If another app is using the microphone (Teams, Zoom & co.), Voxista offers to transcribe your side of the conversation — your audio lands as a take. Transcript and AI minutes go to the clipboard on request; the automatic copy can be switched off separately for each.
A floating pill shows live status: real-time waveform, word counter "✓ 128 words inserted", 10-second countdown on silence.
The same binary as a command-line tool — for Shortcuts, the URL scheme voxista:// (token-protected) and an optional webhook (e.g. n8n).
One click on “Install version …” and Voxista downloads the package, verifies its signature and restarts — no Gatekeeper prompt on updates. A small badge on the menu bar icon shows when one is waiting.
Four steps from hotkey to text in the active app — entirely on your Mac.
Hold or double-tap ⌥ — Voxista records at 16 kHz WAV. The menu bar icon pulses red. Esc cancels.
Whisper, Parakeet or Apple Speech — all run entirely on your hardware. Long recordings are split at silence points and transcribed with progress updates.
Ollama (qwen2.5) cleans up, formats, or translates the raw text — depending on the selected mode. Dictionary and vocabulary are applied.
Text is pasted via clipboard + ⌘V exactly where the cursor is. Optionally followed by an automatic Return.
Voxista as a building block in your automation stack — Shortcuts, shell scripts, webhooks.
# ./build.sh --install sets up the command automatically voxista --transcribe interview.mp3 # audio → text (local) voxista --enhance "um the meeting is tomorrow" --mode clean voxista --list-modes # list available modes
voxista:// — trigger actions directly from other apps, Shortcuts and scripts. Starting a dictation (start, toggle, mode) takes your personal token from Settings → Automation, which also holds a ready-made example URL; stop and cancel work without one.
Each completed dictation is optionally sent to a webhook (e.g. n8n) — secured with a Bearer token stored in the macOS Keychain.
"Run shell script" action — uses the same engine and settings as the running app.
The fastest path is the prebuilt download. If you prefer, build from source instead — both options are described below.
# 1. Download & unzip, drag Voxista.app to /Applications: # → codeberg.org/kleinschmidt/Voxista/releases/latest # 2. First launch: right-click the app → "Open" → "Open" # (Voxista is self-signed, not notarized — one-time prompt only) # Alternatively, in Terminal: xattr -dr com.apple.quarantine /Applications/Voxista.app # 3. The Whisper model is downloaded automatically on first launch (guided setup)
whisper is already bundled in the download — Homebrew is not required. From then on the app updates itself: it checks daily, and one click downloads the new version, verifies its signature and restarts — without a Gatekeeper prompt.
# Homebrew (brew.sh) and Swift Command Line Tools are required # Full Xcode is not needed # Speech recognition brew install whisper-cpp mkdir -p ~/whisper-models curl -L -o ~/whisper-models/ggml-large-v3-turbo-q5_0.bin \ https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo-q5_0.bin # Optional LLM enhancement (local) brew install --cask ollama-app ollama pull qwen2.5:7b
git clone https://codeberg.org/kleinschmidt/Voxista cd Voxista ./setup-signing.sh # once: create a stable signing certificate ./make-icon.sh # generate app icon ./build.sh --install # build and install to /Applications # The built app runs without Homebrew (whisper is bundled)
xattr -dr com.apple.quarantine /Applications/Voxista.app
Or: right-click the app → "Open" → "Open".
What each version brought — newest first.
voxista:// now needs a token for start, toggle and mode — you find it under Settings → Automation, with a ready-made example URL. Without it any website could start a recording and have the transcript pasted into its own window. stop and cancel keep working without one--json/--stats/--export-history, URL scheme voxista://, optional webhook (n8n)Voxista is open source (AGPL v3) and comes without a subscription, without an account, without hidden costs. If it helps you day to day, consider supporting development via Liberapay — completely voluntary and anonymous.
External link — Liberapay has its own privacy policy.