VoiceTyper

VoiceTyper is a small Windows PowerShell dictation app. It records your voice with a hotkey, transcribes the completed recording locally with whisper.cpp, then inserts the final text into the active app.

It was built for cases where real-time Whisper streaming is too delayed or unstable. Instead of typing partial text while you speak, VoiceTyper uses a push-to-talk flow:

1. Press F8 or click Record.
2. Speak.
3. Press F8 again or click Stop.
4. VoiceTyper transcribes the saved WAV file.
5. The transcript is pasted or typed into the target app.

The window is intentionally small, about 300 by 300 pixels, so it can sit beside Codex, Chrome, Notepad, or any other app.

Features

• PowerShell-only app UI using WinForms.
• Local speech-to-text through whisper.cpp and whisper-cli.exe.
• Push-to-talk recording with F8.
• Optional manual Record and Stop buttons.
• Direct 16 kHz mono WAV recording through Windows winmm.dll.
• Active-window targeting, with optional locked target mode.
• Fast clipboard paste mode, with key-by-key typing fallback.
• Feedback sounds when recording starts and stops.
• Configurable model path, hotkey, output method, sounds, and target behavior.
• Logs and recordings saved locally for debugging.

Files

• VoiceTyper.ps1 - the main PowerShell app.
• Install-VoiceTyperBackend.ps1 - setup helper for model and whisper.cpp binaries.
• VoiceTyper.config.example.json - sample config.
• VoiceTyper.config.json - local runtime config, created/updated on first run.
• SPEC.md - detailed implementation specification.

Generated folders:

• models - Whisper GGML model files.
• vendor - local whisper.cpp, SDL2, and build outputs.
• recordings - saved push-to-talk WAV recordings.
• logs - app and transcription logs.

Requirements

• Windows.
• Windows PowerShell 5.1 or PowerShell 7.
• A working microphone allowed by Windows privacy settings.
• whisper-cli.exe from whisper.cpp.
• A local GGML Whisper model, for example ggml-base.en.bin.

The setup helper can download the model and build whisper.cpp locally if the required binaries are missing.

Setup

Run the setup helper:

powershell -ExecutionPolicy Bypass -File .\Install-VoiceTyperBackend.ps1

The helper:

• downloads ggml-base.en.bin if needed;
• looks for existing whisper-stream.exe and whisper-cli.exe;
• downloads SDL2 development files;
• clones/builds whisper.cpp when a suitable prebuilt binary is not available;
• writes discovered paths into VoiceTyper.config.json.

Run

powershell -ExecutionPolicy Bypass -File .\VoiceTyper.ps1

Basic Workflow

1. Start VoiceTyper.
2. Click into the target text field, such as a Codex prompt, Chrome URL bar, Notepad, or an editor.
3. Press F8 or click Record.
4. Speak.
5. Press F8 again or click Stop.
6. Wait for transcription.
7. The final text is inserted into the target app.

The Paste checkbox controls output speed:

• checked: insert the transcript through the clipboard in one operation;
• unchecked: type through SendKeys character by character.

The Lock button is optional. Use it when you want VoiceTyper to keep sending output to one specific window instead of following the active foreground window.

Config

Important config values in VoiceTyper.config.json:

• WhisperCliPath: path to whisper-cli.exe.
• ModelPath: path to the GGML Whisper model.
• Language: spoken language, default en.
• Threads: CPU threads used by Whisper.
• Hotkey: default F8.
• AutoTypeAfterTranscribe: insert final transcript automatically.
• OutputMethod: Paste or Type.
• EnableFeedbackSounds: play sounds on record start/stop.
• RecordStartSound: default Asterisk.
• RecordStopSound: default Exclamation.
• TargetMode: ActiveWindow, Codex, or Locked.
• PreferredTargetTitle: default Codex.
• RefocusTargetBeforeTyping: whether to force focus back to the target before output.

Legacy streaming settings such as WhisperStreamPath, StepMs, LengthMs, KeepMs, MaxTokens, and AudioContext are still present because earlier versions used whisper-stream.exe. The current default flow uses push-to-talk recording plus whisper-cli.exe.

Diagnostics

VoiceTyper writes logs under .\logs:

• VoiceTyper.log - app lifecycle, target selection, recording, transcription, output, and exceptions.
• transcribe-*.stdout.log - whisper-cli.exe transcript output.
• transcribe-*.stderr.log - Whisper model loading and diagnostic output.

Recordings are saved under .\recordings.

Known Limits

• This is not true real-time dictation. It transcribes after you stop recording.
• Clipboard paste mode temporarily replaces the clipboard, sends Ctrl+V, then tries to restore the previous clipboard text.
• Global hotkey handling is polling-based with GetAsyncKeyState, not a registered Windows hotkey.
• The app targets normal desktop text fields. Some elevated/admin apps may not accept pasted or typed input from a non-elevated PowerShell process.
• Accuracy and speed depend on the Whisper model and CPU.

Why Push-To-Talk Instead Of Streaming?

The first implementation used whisper-stream.exe to emit partial text while speaking. On this Windows setup, the stream output behaved more like delayed chunks than reliable real-time dictation. It also revised partial text, which made target insertion harder.

The current design records first, transcribes once, and inserts the final result. This is slower than ideal live dictation, but more predictable and easier to share.