Skip to main content

Accuracy & speed

What affects dictation accuracy, and what Voice Type actually does about it

This page explains the engineering tradeoffs behind dictation quality on Mac: audio conditioning, segmentation, local processing, and the failure modes we are trying to avoid.

Start with the real problem

Dictation quality is not one number. People feel failure when text arrives too late, when punctuation drifts, when background noise pollutes the input, or when the last sentence falls apart after they stop speaking. A serious product needs to deal with all of that, not just brag about a model name.

Why local processing changes the feel

  • Streaming windows: Voice Type transcribes in ~30 second blocks so the interface stays responsive even during long sessions.
  • Quick finalisation: When you release the hotkey, only the last block remains. The point is to keep the stop-to-text loop short enough that dictation still feels usable in the middle of real writing.
  • No uploads: With everything on-device there is no waiting on hotel Wi-Fi or network round-trips before text can land.

What happens before decoding

  • RNNoise VAD: Removes silence and suppresses background noise before words ever reach the recogniser.
  • Signal conditioning: −14 LUFS normalisation, 50 Hz high-pass filtering, and band-limited resampling align the input more closely with the speech distribution the recogniser expects.
  • No heavy prompt hacks: We avoid over-corrective LLM prompts that can make transcripts look confident but drift from what you actually said.

What this does not magically fix

  • Bad microphones still matter: signal conditioning helps, but it cannot rescue terrible source audio forever.
  • Category mistakes still matter: no dictation app replaces Voice Control when the real need is hands-free command of the Mac.
  • Workflow fit still matters: even accurate dictation feels wrong if the text appears in the wrong place or too late for the way you work.

Why users can trust the claims

  • App Store sandboxed: Voice Type is distributed through Apple with notarisation and sandboxing enabled.
  • Authentic reviews: Our App Store listing surfaces real feedback, positive and critical, to keep us accountable.
  • Optional rewrites only: Dictation never hits the network unless you opt into bring-your-own-key rewriting.