Is on-device dictation cheaper than cloud dictation?

For people who dictate regularly, on-device dictation is usually cheaper in cash, because it has no per-minute bill. Its cost moves to your computer, in memory, storage and battery. For very light use, a free cloud tier can cost nothing at all. Renting a server to run a model yourself is usually the most expensive option for one person, because the server bills whether or not anyone is speaking.

The prices below are list prices that we checked on October 16 and 17, 2026. Promotions, volume discounts and taxes change what you pay.

On-device dictation has no per-minute bill, but it uses your hardware

On-device dictation runs the speech model on your own computer, so no provider bills you for each minute. Apple says that on a Mac, voice inputs and transcripts for general text Dictation are processed on your device and not sent to Siri servers ¹.

Open models cost you hardware instead. The large Whisper model in whisper.cpp needs about 3.9 GB of memory and takes 2.9 GiB on disk ². Quantized versions need less memory and disk space, so a smaller model asks less of your machine. Those costs are paid once, in the computer you already own, and they do not grow with each minute you dictate.

Cloud dictation charges a fixed amount for each minute you speak

Cloud dictation has a bill that grows with use. At list prices, AssemblyAI's Universal-2 model costs $0.15 per hour of audio ³. Google's standard rate for its first 500,000 minutes each month works out to $0.96 per hour of audio ⁴. Those two figures are a wide range, and this page does not survey other providers.

A recurring cloud bill never stops while you keep dictating, and on-device dictation avoids it entirely. That is why on-device dictation is usually the cheaper choice for regular daily use, provided your computer has room for the model.

Free cloud tiers beat both options for very light use

Free allowances change the answer for light users. Microsoft's Azure AI Speech free tier includes 5 audio hours a month of standard real-time transcription ⁵. If you dictate a few hours a month, that allowance costs nothing in cash. Free tiers are limited, though. Azure's free allowance is capped at 5 audio hours a month ⁵, so check the terms before you rely on one.

Self-hosting on a rented GPU costs more than cloud for one person

Running your own model on a rented server is the most expensive route for one person. A g4dn.xlarge instance with one NVIDIA T4 GPU starts at $0.526 per hour on demand in the US East region ⁶. The server bills for every hour it is switched on, so an idle server still costs money. Left on all month, around the clock, it costs about $383.98 ⁷. For our 42-hour user, that works out to about $9.14 per hour of dictated audio ⁸, many times the cloud rates above.

A rented server would have to process about 2,560 hours of audio a month to match AssemblyAI's $0.15 per hour ⁹. That is far more than one person dictates. Self-hosting can make sense for a team that shares the server, or for private work that must not leave the company's network, but the price is not the reason.

It depends on your volume, your privacy needs and your hardware

The right choice depends on three things. The table below shows how they usually combine.

Hardware is the one cost that on-device dictation cannot avoid. The large Whisper model needs about 3.9 GB of memory ², so check that your machine has room for the model you want before you rely on it. Battery use is a separate cost, which our question on battery use covers.

What to do: compare your minutes at each rate

1. Measure your month. Log one normal week of dictation and multiply it to a monthly total.
2. Price the cloud. Multiply your monthly hours by the provider's rate per hour, then compare with the free tier's allowance.
3. Check your hardware. Our guide to running Whisper locally explains what a model needs, and the free dictation software guide lists tools that run on your own device.
4. Decide on privacy separately. If the audio must stay on your computer, choose on-device whatever the price.

Yap, which we make, runs its dictation on the device on Mac and Windows. You can see its current plans on our pricing page, and read the download page for system requirements.

This question sits beside why AI dictation apps charge a subscription, and both belong to our AI dictation cost guide within the voice and AI overview. For the terms used here, see on-device speech recognition and cloud dictation in our glossary.