Skip to content
On-Device AI Engine

How WhisperType transcribes speech entirely on your Mac.

Unlike cloud speech recognition services that stream your voice recording across the web to remote servers, WhisperType transcribes speech entirely on your Mac. Using the Whisper speech model compiled to Apple CoreML, it processes audio directly on your Apple Silicon Neural Engine with zero cloud egress.

🎙️ In-Memory Audio CaptureRAM Only

When you hold or tap ⌥Space, AVAudioEngine buffers 16 kHz mono PCM audio directly into memory. Audio is never written to disk or recorded to temporary files, ensuring complete security.

⚡ Apple Neural Engine InferenceCoreML

Audio frames are passed straight to WhisperKit CoreML running on your M-series Neural Engine and GPU. Transcriptions finish in milliseconds with near-zero battery drain.

⌨️ Seamless Text InsertionAccessibility

WhisperType synthesizes a native paste keystroke (⌘V) into whichever app is currently focused. It saves and restores your prior clipboard contents so your copied text is never lost.

🗂️ 100-Item Dictation HistoryQuick Copy

Misplaced a dictated note? Open the Studio dashboard to view your last 100 transcriptions. Each entry includes timestamps and a one-click copy button to retrieve text anytime.

🧠 Memory ManagementRAM Control

Keep your active model pinned in unified memory for instant hotkey response, or use the one-click Unload Model option in the Control Panel to free system RAM instantly when not dictating.

⏱️ Smart Silence DetectionAuto-Stop

Configure auto-stop between 1 and 10 seconds of silence. WhisperType detects when you stop talking and automatically finalizes and inserts your words without requiring a keypress.

✂️ Silence TrimmingPre-Inference

Leading and trailing silence is cut from the buffer before Whisper sees it. This is a correctness fix, not an optimisation: given a long tail of quiet, Whisper hallucinates text to fill it — stray phrases, subtitle credits, sometimes in the wrong language. Removing the silence removes the cause.

🎯 Floating Mic ButtonAlways On Top

An optional round mic button rendered in an NSPanel above every app and on every Space. Click to start and stop, drag to reposition — the position persists across launches. It never steals focus from the app you are typing in.

🧠 On-Device AI CleanupmacOS 26+

Apple's FoundationModels framework rewrites the raw transcription into finished prose using the model already built into macOS. Weak-linked, so older systems simply do not show the option — and like everything else here, it runs on your Mac with zero network egress.

🗂️ Per-App ModesContext Aware

WhisperType records which app was frontmost when recording began and matches it against your rules by bundle identifier. Dictating into Mail produces an email; into Terminal, literal text with nothing added. Four presets ship, and none of them fire until you add one.

Model Catalogue & Performance
Tiny.en (~75 MB)
Ultra-fast, lowest memory usage. Ideal for quick notes and older Apple Silicon Macs with 8 GB RAM.
Base.en (~145 MB)
Fast and highly reliable for everyday dictation, messaging, and emails.
Small.en (~480 MB)
Default & Recommended. Exceptional balance of accuracy, vocabulary recognition, and speed.
Medium.en (~1.5 GB)
Very high accuracy for specialized technical vocabulary, long-form dictation, and accents.
Large v3 Turbo (~3.0 GB)
Multilingual. Distilled from Large v3 — close to its accuracy at a fraction of the inference cost. The one to pick if you dictate in more than one language.
Large v3 (~3 GB)
Multilingual. The flagship, the checkpoint behind most published speech-research benchmarks. The most accurate option and the slowest.
First Load vs. Every Load After
Cold load
CoreML compiles the model for the Neural Engine on first use and caches the result. Measured on an M4 Mac: about 16s for Medium, about 45s for Large v3 Turbo, whose audio encoder now compiles for the GPU instead of the Neural Engine. It happens once per model.
Warm load
0.4s and 0.9s respectively — a 9× and 103× speed-up. The progress bar is calibrated on the slow case, so it paces a first load honestly instead of stalling at 95%.
Download vs. load
The two phases are reported separately — Downloading model files (n%) and Loading into memory (n%) — so a percentage always tells you which process it belongs to.
What It Costs
14-day trial
The whole app, unlimited, no account and no card.
Free plan, forever
10 minutes of dictation a day, every feature still enabled. The meter counts transcribed audio after silence trimming — not wall clock — so an app left open all day costs nothing, and a cancelled dictation is free. Resets at local midnight.
Pro
$24.99 once removes the daily limit on up to three Macs. No subscription, all future updates included.
Complete Privacy Guarantee: An internet connection is required only once to download your selected CoreML model from HuggingFace. After download, WhisperType works completely offline with your Wi-Fi disconnected.