WhisperTyper

Hold a hotkey, speak, release — your words appear at the cursor. Local, private, GPU-accelerated.

WhisperTyper runs OpenAI Whisper entirely on your machine using Const-me's DirectCompute implementation, with no cloud, no subscription, and no data leaving your PC. It supports NVIDIA, AMD, and Intel GPUs out of the box via DirectCompute (Direct3D 11).

Features

• Hold-to-record — hold a configurable hotkey while speaking, release to transcribe
• Types at cursor — output appears wherever your cursor is, in any app
• GPU-accelerated — runs on NVIDIA, AMD, and Intel GPUs via DirectCompute; falls back to CPU automatically
• Local & private — no internet connection required after setup
• Per-app profiles — automatically switch model, language, and LLM prompt based on whichever application is in focus
• LLM post-processing — route transcribed text through a local or cloud LLM before it's typed; configure a different prompt per app
• Sentence streaming — complete sentences are typed at the cursor mid-recording, so you don't wait until the end (disabled automatically when LLM post-processing is active)
• GGML model support — works with any Whisper GGML .bin model (base, small, medium, large, turbo)
• Model manager — browse, download, and switch models from within the app
• History log — searchable record of past transcriptions; copy or re-type any entry
• Custom dictionary — fix words Whisper consistently mishears (e.g. "eye phone" → "iPhone")
• Filler word removal — automatically strip "um", "uh", "like", etc. before typing
• Modern dark UI — system tray, minimize to background, settings persist between sessions

Per-app profiles and LLM post-processing

Each application on your system can have its own profile. When you dictate into a focused app, WhisperTyper applies that profile's settings — including sending your transcribed words through a configurable LLM prompt before typing them. This enables things like:

• Formal tone — speak casually, have it typed as polished professional language in Outlook or Teams
• Console commands — describe what you want in plain language; the LLM turns it into the actual shell command
• Seamless translation — speak in your native language and have the output typed in another, with tone already set in the prompt
• App-specific rewrites — Slack profile uses casual language, email client uses formal, each applied automatically based on the focused window

Requirements

• Windows 10 or 11 (x64)
• .NET 10 Runtime (or SDK to build from source)
• A GPU with DirectCompute support (Direct3D 11) — or CPU fallback
• A Whisper GGML model file (.bin) — download one from the built-in Model Manager or see below

Downloading a model

WhisperTyper does not bundle a model. Use the built-in Models panel to download one, or grab it manually from Hugging Face:

Model	Size	Notes
ggml-base.bin	142 MB	Fast, decent accuracy
ggml-small.bin	466 MB	Good balance
ggml-large-v3-turbo.bin	1.5 GB	Best accuracy, recommended

Building from source

git clone https://github.com/joyware/WhisperTyper.git
cd WhisperTyper
dotnet build -c Release

The native Whisper.dll (DirectCompute inference engine) is downloaded automatically via the WhisperNet NuGet package on first build.

Usage

1. Launch WhisperTyper
2. Select your model file (or download one from the Models panel), GPU, and microphone
3. Wait for the status indicator to turn green ("Ready")
4. Click into any text field in any application
5. Hold your hotkey (default: Caps Lock), speak, then release
6. Text appears at your cursor; sentences stream in mid-recording when possible

WhisperTyper minimizes to the system tray and stays active in the background.

Hotkey options

Any key or mouse button can be configured as the hotkey using the built-in key recorder. Commonly used options:

Key / Button	Notes
Caps Lock	Default; Caps Lock toggle is suppressed while in use
Scroll Lock	Good alternative if Caps Lock is needed
Mouse Button 4 / 5	Thumb buttons on most mice — recommended for frequent use
Middle Mouse Button	Works well if not used for other purposes
Left Alt	Suppresses Alt menu activation
Left Ctrl
F9 / F10	Function keys

Modifier combinations (e.g. Ctrl+CapsLock) are also supported. Note that sentence streaming is disabled when a modifier key is part of the hotkey, as it interferes with text injection.

How it works

• Audio is captured via WASAPI (NAudio) while the hotkey is held
• Every 4 seconds, a snapshot of the audio is transcribed and any new complete sentences are typed immediately at the cursor
• On release, the remaining audio is transcribed and the untyped remainder is injected
• If a profile has LLM post-processing enabled, the full transcription is sent to the LLM first and the result is typed instead
• Text is injected using the Windows SendInput API with KEYEVENTF_UNICODE — works in any application, supports full Unicode
• Filler words and custom dictionary replacements are applied before any text reaches the cursor

Dependencies

Package	Purpose
WhisperNet 1.12.0	Whisper inference via Const-me's DirectCompute DLL
NAudio 2.2.1	WASAPI audio capture

License

MIT — © 2025 Joyware