How accurate is on-device voice transcription?
Modern on-device transcription is very good. Local Whisper models reach roughly 95%+ accuracy on clear speech, close to cloud services for everyday dictation. Cloud models can still edge ahead on dense technical vocabulary, but for email, messaging, and documents the gap is small, and your audio stays private.
Last updated . Prices and vendor claims come from the sources listed at the end of the page; re-check before purchase.
A few years ago, "local" meant "less accurate." That's largely no longer true. Whisper-based on-device transcription (what Relay, Superwhisper, and VoiceInk use) is accurate enough that most people can't tell the difference from cloud for normal speech.
On-device vs cloud accuracy
Everyday speech: on-device and cloud are roughly equivalent. Both handle natural dictation, punctuation, and common words well.
Dense technical vocabulary: cloud models tuned for jargon (Aqua Voice, for example) can be more accurate on variable names, medical terms, and unusual proper nouns. If your dictation is mostly specialized terminology, that may matter.
The trade-off: cloud accuracy on jargon comes at the cost of sending your audio off-device. For most users, on-device accuracy is more than enough and privacy is the better deal.
What actually affects accuracy
- Microphone: a decent mic matters more than the app. Built-in MacBook mics are fine; a headset or external mic is better in noise.
- Background noise: quiet rooms transcribe far better than cafes.
- Model size: larger Whisper models are more accurate but slower. Most apps pick a sensible default.
- Accent and clarity: speaking at a steady pace helps more than people expect.
- Custom vocabulary: apps that learn your jargon improve on names and technical terms over time.
How to get the best accuracy
Use a good mic in a quiet space, speak naturally at a steady pace, and let the app clean up formatting afterward. Relay's Clean and Rewrite modes fix punctuation and filler on-device, so a slightly imperfect raw transcript still comes out polished. For the accuracy-vs-privacy trade-off in depth, see Relay vs Aqua Voice.
Sources
- Aqua Voice pricing page (checked 2026-09-10)
- Relay release notes (GitHub) (checked 2026-09-10)
Frequently asked questions
How accurate is local Whisper transcription?
On clear speech with a decent microphone, local Whisper models reach roughly 95% or higher accuracy, close to cloud services for everyday dictation. Accuracy drops with background noise, heavy accents, or very specialized vocabulary.
Is cloud dictation more accurate than on-device?
For everyday speech they're roughly equivalent. Cloud models tuned for technical vocabulary (like Aqua Voice) can be more accurate on dense jargon and unusual terms. The trade-off is that cloud apps send your audio to their servers.
Why is my dictation accuracy low?
Usually the microphone or environment, not the app. Background noise, a distant or low-quality mic, speaking too fast, or heavy specialized vocabulary all reduce accuracy. A headset mic in a quiet room makes a big difference.
Does on-device transcription get better over time?
Local models themselves are fixed until you update the app, but apps that support custom vocabulary learn your specific terms and names, which improves accuracy on the words you use most.
Try accurate, private dictation
Relay transcribes on-device with Whisper, then cleans it up. Free during beta.
Download for MacFree during the beta · Apple Silicon (M1 or later) · Your voice never leaves your Mac
Get a license key and release notes by email (optional; the download does not require it)