Which model should you use?

Cloud and on-device are different bargains

Every row on this page is one or the other, and the badge says which. The distinction is not a detail — it changes what you pay, what you wait for, and where your voice ends up.

Where these numbers come from

Accuracy

Word error rate comes from the Artificial Analysis speech-to-text leaderboard — an independent third party, not us. We do not publish an accuracy claim of our own here. A model they have not measured shows a dash and is marked not benchmarked; it scores neutrally rather than being flattered by a number we invented.

Why some on-device models borrow a cloud model's score

Whisper and Parakeet are open weights. The leaderboard measured the same weight files we download, just running on someone else's hardware — so the error rate carries over and is marked same weights. Speed does not carry over, which is why those rows never borrow a speed figure. The remaining local models have no published benchmark at all and fall back to the app's own accuracy rating, marked app rating. Without that fallback the smallest, roughest models would score like frontier ones purely for being free and private.

Cost

Credits per audio minute, read from the same catalog the desktop apps read, so the price here is the price the app charges you. 1,000 credits are $1, which makes a model at 4.5 credits a minute $4.50 per 1,000 minutes of audio. On-device models cost nothing to run, so they score full marks on cost no matter how you weight it.

Speed

Two numbers, and the difference matters. Per audio minute is how long a minute of speech takes to come back, from the leaderboard's published speed factor; on-device rows estimate it from the app's own speed rating, so read those as an ordering, not a promise. That is the number the ranking uses, because it is the only one we hold for every model. Measured clip is our own median from the last 90 days — the same measurements behind the latency page, taken from clips under ten seconds because that is what dictation is. It is a round trip for one clip, not a throughput: most of it is the network and the provider's fixed overhead rather than decoding, and we do not store the clip's length, so it cannot be scaled up to a minute. Feeding it into the ranking anyway would punish the providers we happen to know most about, so it sits in its own column instead.

Privacy

Three steps, by where the audio goes. On-device scores full marks: nothing is transmitted. A model you can use with your own API key scores half: the audio still reaches the vendor, but on your account rather than through us. A cloud-tier-only model scores lowest. Turn the privacy slider up and the ranking moves to on-device models, which is the honest answer to that question.

Your region

HyperWhisper Cloud runs in 17 regions and routes you to the nearest one. We ask a small endpoint which of those you are closest to so the speed numbers are the ones you would actually get; the answer is used once to pick a row and is never stored. You can override it with the dropdown. Off our edge — a local dev server, say — nothing is detected and the busiest region is selected instead.

What this page will not tell you

A ranking is not a recommendation for a job it has never seen. If you dictate medical or legal terms, transcribe heavy accents, or need speaker labels, the model that wins here on paper may still lose on your audio. Every model listed is switchable in Settings → Transcription, so the last word is a minute of your own speech, not a number on a website.