On-device dictation went from unusable to excellent in three years. Here's what runs locally now, what hardware it needs, and what you still give up

How it works

What 'offline' has to cover

A dictation app is only genuinely offline if three things all run locally: the speech recognition, any text cleanup, and the licensing check. The first two are obvious. The third is the gotcha — several apps run their models on-device but phone home to validate a subscription, which means they silently stop working on a plane. When we mark an app as fully on-device in the ranking, we have tested it with the network cable out, not just read the marketing page.

The hardware question

Local dictation trades network for compute. The practical ladder on Apple Silicon: Whisper tiny and base models run on anything and produce mediocre text; small and medium are the sweet spot for M1-and-later machines, transcribing faster than real time; large-v3 is where local quality starts rivalling the cloud, and wants 16 GB of RAM and an M-series chip to stay responsive. Parakeet V3 hits comparable accuracy to large-v3 at a fraction of the memory, which is why it has become the default local engine in apps that support it.

On Windows the picture is spottier. whisper.cpp runs well on modern CPUs and dramatically better with an NVIDIA GPU; without one, the larger models lag. Dragon Professional, the old-school outlier, has run entirely on-device for decades on ordinary Windows hardware — proof that local dictation was always possible, just not good until recently.

The quality gap, honestly stated

For English and the largest European languages, a well-configured local setup in 2026 is within a few accuracy points of the best cloud apps for clean speech — close enough that the difference rarely matters in practice. The gaps that remain: heavy accents and noisy rooms (cloud models are trained on more of both), smaller languages (the long tail of the Whisper list is rough everywhere, worse at small model sizes), and the LLM cleanup pass, where the cloud apps' big rewrite models still outclass anything that fits on a laptop.

The gap also runs the other way on one dimension: latency. A local model starts producing text the moment you stop speaking, with no network round trip. Local dictation feels instant in a way cloud dictation does not.

Which apps are genuinely offline

Fully offline, tested with no network: Handy, OpenWhispr, VoiceInk, MacWhisper's local mode, Superwhisper, SpeechPulse, and Dragon Professional. Offline recognition but cloud-dependent licensing or cleanup: several mid-tier apps we won't name-and-shame here — check our per-review privacy sections, where we note what breaks when the network drops.

Cloud-only: Wispr Flow, Aqua Voice, Willow Voice, Monologue. Excellent products, but they are radios as much as microphones — nothing works without a connection, and no setting changes that.

Frequently asked questions