Inspiration
Voice input is useful, but changing applications, copying text, and restoring focus breaks the user's flow. VoxPilot was designed around a simpler interaction: place the caret anywhere, use one shortcut, speak, and continue working.
What it does
VoxPilot is a compact native Windows voice-to-text companion that transcribes through OpenAI or OpenRouter and types the result into the application that was active when dictation began.
- Selectable OpenAI and OpenRouter transcription providers
- Press-to-toggle and hold-to-talk workflows
- Microphone selection and a local microphone test
- Automatic language detection or a selected language hint
- Exact transcription plus Polished and Notes modes powered by GPT-5.6 Terra
- Immediate Esc cancellation that discards audio without contacting a provider
- Unicode text injection into the previously focused application
- Secure per-provider keys in Windows Credential Manager
- Memory-only audio capture, tray recovery, configurable shortcuts, and clipboard fallback
How we built it
VoxPilot is a native .NET 10 WPF application written in C#. NAudio captures WAV audio in memory. Dedicated clients handle OpenAI and OpenRouter transcription. Smart-text modes use the OpenAI Responses API with GPT-5.6 Terra and an exact-transcript fallback. Windows APIs provide global hotkeys, key-release detection, foreground-window tracking, Unicode input injection, notification-area behavior, and secure credential storage.
Codex with GPT-5.6 was the primary engineering collaborator. It helped scaffold the WPF architecture, implement and debug Windows integrations, add provider support, refine the interface through rendered previews, verify behavior, and package the self-contained release. The entrant directed the product scope, native-Windows architecture, interaction model, privacy boundaries, provider strategy, and final design decisions.
Challenges
The hardest part was reliably returning text to the correct destination without letting the floating interface or dictation widget steal focus. Windows also blocks non-elevated applications from injecting input into elevated windows, so VoxPilot explains that limitation and falls back to the clipboard when direct typing is blocked.
Another challenge was making recording status visible while the main window was hidden. The separate waveform widget is deliberately non-activating, allowing users to see progress without losing their caret position.
Accomplishments
- A complete native product experience rather than a transcription proof of concept
- Reliable global-shortcut dictation across ordinary Windows applications
- Memory-only recording and secure provider-specific credential storage
- Push-to-talk, safe cancellation, microphone testing, and GPT-5.6 smart-text modes
- A self-contained Windows release that judges can run without installing the .NET SDK
What's next
- Optional local/offline transcription
- User-defined vocabulary and additional formatting profiles
- Streaming transcription for lower perceived latency
- A signed installer and automatic updates
