TypeWhisper for macOS

Problem / Context

Speech-to-text is most useful when it disappears into the operating system, but voice data is also unusually sensitive. TypeWhisper approaches that tension directly: make private, on-device transcription the default and keep cloud processing optional.

I contribute to the public macOS project and follow its work at the boundary between native interaction, model execution, and clear user control.

Constraints

• Maintain a native macOS experience instead of a web interface wrapped in a window.
• Keep the private path genuinely local.
• Make model and provider choices understandable to people who do not want to manage ML infrastructure.
• Preserve fast keyboard-driven use across other applications.

Approach and key decisions

The project uses Swift and the native platform surface so recording, permissions, shortcuts, and text insertion behave like part of macOS. The privacy model is visible in product decisions: local models remain useful without an account, while optional providers are explicit.

Contribution work in a mature public repository also means respecting existing architecture and review conventions. Small, legible changes beat broad rewrites.

Outcome

TypeWhisper has become a widely used open-source macOS project with an active contributor community. For me, the valuable outcome is practical: helping a privacy-sensitive AI workflow feel like a normal, dependable desktop tool.