Every dictation app turns useAuthSession into "use auth session". VaniQ reads the repo you're working in first, so your own identifiers beat the phonetically similar English. Hold a key, talk, let go. 190ms from key-up to text, all on the Neural Engine. No server, no telemetry, works with the Wi-Fi off. Wispr Flow sends your audio to its servers. Superwhisper and MacWhisper run locally but use general models, so they mangle the same names. Reading your repo is the difference.
Hey everyone,
I talk to Claude Code and Cursor most of the day, and every dictation app I tried broke on the same thing. I'd say "useAuthSession" and get "use auth session". "kCGEventTap" came out as "KCG event tap". Fixing the names afterwards took longer than just typing them, so I kept going back to the keyboard.
VaniQ reads the repo before it listens. It pulls the symbols out of whatever project has focus and biases decoding toward them, so your own identifiers beat the phonetically similar English. Switch projects and it re-indexes on focus change, so the vocabulary is already right by the time you start talking.
Everything runs on-device on the Neural Engine. No server, no telemetry, and it keeps working with the Wi-Fi off after the first model download. It also tags its clipboard writes so Raycast and Alfred skip them, which I only thought to check after realising my clipboard manager had quietly built a searchable log of everything I'd ever dictated.
Numbers I measured rather than guessed: 190ms from letting go of the key to text on screen, 98.4% word accuracy overall, 97.2% on speech deliberately stuffed with code names.
Why this over the other voice apps: Wispr Flow sends your audio to their servers. Superwhisper and MacWhisper can run locally but use general-purpose models, so they hit the same identifier problem I started with. Reading your repo is the difference.
Limits up front: Apple Silicon only, macOS 14+, English only, $9.99/mo and no free trial yet. I know that last one is a real ask for something you can't poke at first, and it's what I'm working on next.
Would love feedback, especially on what else should feed the vocabulary beyond the repo.