A small key. A complete speech system.

Hold F9, speak, and release. Quill records your voice, sends it to a server for transcription, and attempts to paste the words where you were typing.

The internal Windows tool connects native controls, local speech models, and the receiving application.

Start with usable audio

The Tauri, TypeScript, and Rust client moves recording and network work off the keyboard hook, keeping that callback responsive.

An accidental F9 tap shorter than 0.35 seconds is rejected before upload. The client also checks for an effectively silent microphone. These filters address unsuitable input before asking a model to interpret it.

Preserve the words when cleanup fails

The server uses faster-whisper large-v3 for transcription. An optional Ollama model then adds punctuation and removes filler.

An editor might answer a dictated question instead of preserving it. Quill checks output length and word overlap, returning the original transcription when cleanup fails or strays too far. These checks cannot guarantee that every word retains its meaning.

Keep the text when paste cannot proceed

Quill writes the transcript to the clipboard before attempting a paste. It checks whether the foreground window changed, waits for held modifier keys, and checks whether the destination has higher Windows privileges that could block input.

The clipboard provides a manual fallback, replacing its previous contents. Quill cannot confirm that every destination accepted the paste; delivery needs its own checks.

Share the hardware deliberately

Short dictation benefits from a ready model. Long recordings need a queue that survives interruption. The dictation backend can warm its model and release it after inactivity; the lecture transcriber checks available GPU memory and uses a separate process for each job.

Quill needs its server connection and human review. It demonstrates complete product integration: native controls, models, shared resources, and recoverable text delivery.