Verbal

Verbal is a Windows desktop app that transcribes your speech locally and can type the transcript into a text field automatically or allow you to copy the transcribed text. Whisper model files are not included.

Using Verbal

Requirements

• Windows 10 or later
• A microphone
• At least one compatible Whisper ggml-*.bin model in whisper_cpp\models

The GitHub release is self-contained and includes the app runtime and CPU Whisper CLI. You do not need to install .NET to use that release.

Install a Whisper model

Download a compatible Whisper model in ggml-*.bin format and place it in the whisper_cpp\models folder next to Verbal.exe. Model weights can be quite large and are intentionally not included with Verbal. The app looks for models when it starts, so restart it after adding or removing model files.

For example, ggml-base.en.bin is a suitable English model. Other compatible models can be obtained from the whisper.cpp model collection.

I recommend starting with the ggml-base.en.bin model. It is a good balance between size and accuracy. If you find you need more accuracy, try a larger one.

Record and insert a transcript

1. Open the destination app and select the text field where you want the transcript to go.
2. Start recording in Verbal:
   • Hold F8 by default, or Hold to talk, to record while pressed. Release to stop and transcribe.
   • Press F9 by default, or Start recording, to toggle recording on and off.
3. The transcript appears in Verbal's editable text area. Automatic typing is enabled by default. Turn it off if you want to review or edit the transcript before inserting it.
4. Use Insert into target to type the transcript into the selected field, or Copy to put it on the clipboard.

Use Settings to select your model and microphone, and to change either global keyboard shortcut. Click a shortcut button and press the desired key or combination. Function keys work on their own; letter and number keys require a modifier such as Ctrl, Alt, Shift, or Windows. The defaults are F8 for push-to-talk and F9 for toggle recording. These shortcuts work while Verbal is in the background or minimized. Choose combinations that do not conflict with shortcuts used by other apps. Minimize Verbal to move it into the Windows notification area. The tray icon shows the live microphone level while recording. Double-click the icon to reopen the window, or right-click it to open Verbal, start or stop recording, or exit. Turn on Start with Windows in Settings to launch Verbal automatically when you sign in; it starts minimized in the notification area. Turn the option off to remove it from Windows startup.

Verbal remembers the last active window outside itself and restores it before typing. Windows may prevent foreground activation in some situations. If that happens, the transcript remains in Verbal so you can select the destination field and insert it again.

Audio is recorded as temporary 16 kHz mono PCM WAV files and deleted after transcription. Recording and transcription run locally on your computer.

Download and use a GitHub release

1. Open Verbal Releases.
2. Download the appropriate release archive for Windows and extract it.
3. Keep the extracted folder together. It contains Verbal.exe and the whisper_cpp folder with the Whisper CLI and its required DLLs.
4. Add at least one compatible model file to whisper_cpp\models (see Install a Whisper model).
5. Run Verbal.exe.

Do not move Verbal.exe out of the extracted folder; the app needs the adjacent Whisper files. The release does not contain model weights, so add a model before recording.

NVIDIA CUDA Support

1. Download cuda-inference.zip from releases.
2. Extract the contents of the ZIP to the root folder of Verbal.
3. Run Verbal and check the settings to set your inference preference.

See Models and Inference below for more info.

Clone the repository and build from source

Requirements

• Windows 10 or later
• .NET 8 SDK
• A microphone
• At least one compatible Whisper ggml-*.bin model in whisper_cpp\models to transcribe speech

The repository includes the CPU whisper-cli.exe and runtime DLLs. Model weights are excluded from the repository and must be added separately.

Clone the repository:

git clone https://github.com/Aftonone/verbal.git
cd verbal

Add a compatible model to whisper_cpp\models if you want to transcribe recordings, then run the app from source:

dotnet run

Build a self-contained release

Publish a Windows x64 release with the .NET runtime included:

dotnet publish .\Verbal.csproj -c Release -r win-x64 --self-contained true -o .\publish

Run .\publish\Verbal.exe and keep the entire publish folder together. The publish includes the CPU Whisper CLI and its DLLs directly in publish\whisper_cpp. If whisper_cpp\cuda\whisper-cli.exe and its DLLs are present, those files are also included under publish\whisper_cpp\cuda. Model weights are not included; add models to publish\whisper_cpp\models. To prepare a clean release, publish to an empty output folder or remove an older output folder first so stale files are not left behind.

Models and inference

Verbal uses CPU inference by default for compatibility across Windows systems. When whisper_cpp\cuda\whisper-cli.exe is present, an inference selector appears in Settings so users can choose CPU or CUDA. CUDA inference requires compatible NVIDIA hardware and drivers. Model files are user-managed: add or remove compatible ggml-*.bin files in whisper_cpp\models and restart the app to refresh the model selector. Preferences for the inference backend, microphone, model, and appearance are stored locally for the current Windows user.

See NVIDIA CUDA Installation Guide for help on setting up CUDA.