Local vs cloud dictation: what actually differs
Most comparisons argue about accuracy. That is the least interesting difference, and the one you can measure yourself in ten minutes.
Accuracy: cloud wins, by less than you think
A large cloud model running on a datacenter GPU has an advantage: it can be 20 times bigger than a model that has to fit in your laptop's memory. On clean, well-recorded speech the difference shows up as a couple of percent of words.
Where it shows up more: heavy accents, unusual names, crosstalk, and bad microphones. Where it shows up less: the twelve words you actually dictate most often, especially once the app has a vocabulary list for them.
The practical test is to dictate a paragraph you would really send, three times, on both. One of them will make a mistake you have to fix. That is the whole comparison.
Latency: local wins, and it is not close
Cloud involves a round trip: upload the audio, wait for the model, download the text. On a good connection that is 400–900 ms after you stop talking; on hotel Wi-Fi it is unpredictable.
Local has no round trip. On Apple Silicon a well-chosen model decodes a ten-second dictation in under a second, and can start producing text while you are still speaking.
Privacy: only one of them can be checked
Turn off the network and dictate. That is a complete test of a local app, and no test exists for a cloud one — you are trusting a policy.
Price: the shapes are different
Cloud costs the vendor money per minute, so it is priced as a subscription and it has to be. Local costs the vendor nothing after the download, so it can be sold once. A $15/month subscription is $540 over three years; a $39 one-time purchase is $39.
Where cloud is the right answer
If you dictate hour-long interviews, transcribe noisy field recordings, or need the lowest error rate available and have nothing confidential in the audio, a cloud service is the better tool. Say so plainly: local-first is a trade, not a free lunch.
Where local is the right answer
If the words are private — health, legal, HR, code under NDA, messages to family — or if you work on planes, trains, and flaky hotel Wi-Fi, uploading is either a risk or an inconvenience you do not need.
Frequently asked
Can a local app use a cloud model too? Yes, as an opt-in. That is the sensible design: local by default, cloud only when you ask for it, with the boundary visible.
Does a local model need a big download? One model, one time. Inmost downloads nothing until you choose an engine.
Is local slower to start? The first dictation after a reboot loads the model into memory; after that it is resident.
Inmost runs speech recognition on your Mac, measures the models against your own voice, and is $39 once. See the numbers.