AI Dictation Cost: Cloud vs On-Device Economics

Cloud dictation has a per-minute bill, and each minute is cheap. At list prices, speech recognition from a major provider costs about a quarter of a cent to 1.6 cents per minute ^{1 2}. A cleanup pass by a language model adds less than a cent per hour when it runs on a small model ³. The bill grows with use. An illustrative user who dictates two hours on each of 21 workdays sends 2,520 minutes a month to the cloud ⁴, which costs about 6 to 40 dollars a month at the providers we checked ^{5 6}. That gap is one reason to look at subscription prices, which we cover below. On-device dictation has no per-minute bill. Its cost is paid by your computer, in memory, storage and battery.

The prices below are list prices that we checked on October 16 and 17, 2026. Promotions, volume discounts, taxes and plan changes all alter what you actually pay.

Cloud speech recognition costs about a quarter of a cent to 1.6 cents a minute

A cloud dictation app sends your audio to a speech model that charges for every minute it processes. The price depends heavily on the provider. AssemblyAI lists its Universal-2 model at $0.15 per hour of audio ¹, which works out to a quarter of a cent a minute. OpenAI lists gpt-4o-mini-transcribe at $0.003 per minute ⁷ and gpt-4o-transcribe at $0.006 per minute ⁸. Amazon Transcribe's batch price is $0.006 per minute in its worked example for the US East (N. Virginia) region ⁹. Google's V2 standard recognition costs $0.016 per minute for its first 500,000 minutes each month ².

Deepgram lists its Nova-3 monolingual model at $0.0077 per minute at its regular price, with a current price of $0.0048 per minute ¹⁴. Prices also fall with volume. Google's V2 standard rate drops to $0.004 per minute for accounts that process 2,000,000 minutes or more each month ¹⁵, or $0.24 per hour of audio ¹⁶. Those volumes belong to companies that run transcription at scale, not to one person at a desk.

The table shows only the audio. Google's list separates its rates by product, so check the exact SKU on the page you are billed under before you compare. Google's V1 product charges $0.016 per minute after the first 60 minutes each month when data logging is on ¹⁷, which is the same first-tier rate as V2 but with a different free allowance.

Language-model cleanup adds well under a cent per hour on a small model

Cleanup is the optional second pass that removes fillers, applies your corrections and adds punctuation, as described in our guide to AI cleanup for dictation. It is billed by tokens, the pieces of text a model reads and writes. Anthropic lists Claude Haiku 5.5 at $0.10 per million input tokens for prompts up to 100,000 tokens ¹⁸ and $0.50 per million output tokens ¹⁹. Its larger model, Claude Opus 5.5, is listed at $4 per million input tokens ²⁰ and $20 per million output tokens ²¹.

To turn those prices into a cost per hour, we need the speed of speech. Brysbaert uses 150 words per minute as the speech rate in a theoretical argument that the language system can take in about 300 words per minute ²². We treat 150 as a working figure, not a measured average. At that rate an hour of dictation is 9,000 words. We then assume 1.3 tokens for each English word. That ratio is our assumption, and we could not open an official tokenizer page to confirm it. Under that assumption the cleanup pass reads about 11,700 tokens and writes about 11,700 tokens for each hour of speech ³.

The conclusion does not depend much on the exact token ratio. The cost scales in direct proportion to the ratio, and at 1.3 tokens per word a Haiku cleanup pass costs about 0.7 cents an hour ³. The choice of model matters far more than the choice of ratio. On the largest model, cleanup costs about 28 cents an hour ²³, which is in the same range as transcription at several providers, and on a small model it is a rounding error. The chart puts the transcription and cleanup numbers on one scale.

A heavy dictation habit adds up to about 42 hours of audio a month

We found no published survey that measures how many minutes people spend dictating each day. The nearest figure is about writing rather than speaking. In a 2016 survey of 547 business people who write at least two hours a week for work, respondents averaged 20.4 hours of writing per week, excluding email ²⁴. That figure is about writing, not dictation, so it sets a scale rather than a forecast.

We therefore use an illustrative heavy user: two hours of dictation on each of 21 workdays. That is 2,520 minutes, or 42 hours, of audio a month ⁴. The table prices that one user at three transcription rates.

These figures cover transcription only. They leave out cleanup, storage, the app's own overheads and any taxes. A light user who dictates 30 minutes a day is roughly a quarter of this load, so the same provider charges that user about a quarter of the amount. The usage pattern is our illustration, not a measured average. Anyone can replace it with their own minutes: multiply the monthly minutes by the per-minute price.

Subscriptions cover recurring cloud costs, margins and the heaviest users

A cloud bill recurs every month and grows with each user's minutes, so a cloud dictation app has a recurring cost per user. Its price tends to recur too. Superwhisper lists its Pro plan at $8.49 per month ²⁵, $84.99 per year ²⁶, or $249.99 once as a lifetime purchase ²⁷. The annual plan works out to about $7.08 a month ²⁸, and the lifetime price equals about 29 months of the monthly plan ²⁹. Wispr Flow lists its Pro plan at $12 per user per month ³⁰, and Aqua Voice lists its Pro plan at $10 per month when billed month to month ³¹.

Set those prices against the illustrative user above. The $8.49 monthly plan ²⁵ covers that user's bill at AssemblyAI's rate ⁵, but not at gpt-4o-transcribe's rate ⁴, and it does not come close at Google's standard rate ⁶. A subscription price has to cover the average user and the heavy user together, and any cleanup and margin on top.

A one-time price can leave the vendor paying for the cloud calls a buyer makes long after the purchase, which is one reason subscriptions are common. The lifetime price here equals about 29 months of the monthly price ²⁹. We cannot see how much of each app's work runs in its own cloud, and some apps run models on your device, so the prices show what you pay and not how each app is built. For a longer answer, see why AI dictation apps charge a subscription, and compare the plans in our AI dictation apps compared guide.

On-device dictation has no per-minute cost but uses memory, battery and storage

On-device dictation runs the speech model on your own computer or phone, so it has no per-minute charge from a provider. The costs appear in other places. The whisper.cpp project estimates that the large Whisper model needs about 3.9 GB of memory and takes 2.9 GiB on disk ³². Apple's upgraded Apple Intelligence dictation on iPad is available in English on M4 or later models with at least 12 GB of unified memory ³³. Apple describes a compact on-device model of approximately 3 billion parameters, optimized to run efficiently on Apple silicon, as part of its Apple Intelligence features ³⁴.

Battery use is not measured in the sources we found

We could not find a published measurement of how much battery a dictation session uses on a laptop or phone. Our search for energy data on Whisper found power numbers for embedded boards and project claims, but no measured laptop battery test. The closest rigorous figure concerns a streaming recognizer in a research paper. Li and colleagues report that component-level design changes cut the energy use of an on-device streaming speech recognition model by up to 47 percent while keeping accuracy comparable ³⁵. That is a cut in energy use for one model design, not a battery-life figure, and the model is not Whisper.

So the honest answer is that on-device dictation spends your battery, and the amount depends on the model, the chip and the length of the session. To measure it yourself, run the same dictation session with the app on and off and compare the energy figures your operating system reports. For the question on battery in its own right, see does on-device AI dictation drain your battery. For the trade-off in general, see our guide to on-device and cloud speech recognition, and for a hands-on setup of a local model, the guide to running Whisper locally.

On a Mac, the built-in dictation is one example of the on-device approach. 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 ³⁶. The cost there falls on the hardware you already own, so the Apple silicon generation of the machine matters. Our guide to Apple's privacy for dictation covers what Apple keeps.

Yap, which we make, runs its dictation on the device on Mac and Windows, so its cost model sits in the right-hand column of the table above. It still needs a memory and battery budget, which is the trade we describe here.

Free dictation is paid for by the platform, by your data or by a vendor's other products

Every free dictation tool has someone paying for the speech model. In the built-in case the platform owner pays, usually by running the model on your device. In the third-party case the vendor pays, either by charging you later, through other products, or through the data it collects. We cannot list every app's business model, so the test is practical: read the privacy page before you assume a free app is free.

Two examples show the difference. Apple's own documentation says that on the Mac, general Dictation is processed on the device, but if you opt in to Improve Siri and Dictation, only its own employees, under confidentiality obligations, can access the audio ³⁷. The cost of that opt-in is privacy, not money. Citizen Lab, in its 2024 review of keyboard apps, recommends that at-risk users consider switching from a cloud-based keyboard app to one that works entirely on-device ³⁸. A free cloud keyboard can therefore carry a cost that does not appear on the price list.

Choosing a dictation setup by cost

The numbers point to four choices, and the right one depends on how much you speak.

1. Light use, a few hours a month. At every rate in our table, cloud transcription costs under a dollar an hour ¹³, so a few hours a month costs only a few dollars, and a pay-as-you-go service or a free tier is usually the cheapest route. The free dictation software guide lists the options.
2. Heavy daily use. At two hours a day, the cloud bill ranges from about 6 to 40 dollars a month before cleanup ^{5 6}. At AssemblyAI's rate, pay-as-you-go is cheaper than the $8.49 monthly plan ²⁵, but at Google's standard rate a flat subscription or an on-device tool is the better value, and the lifetime plan pays for itself after about 29 months at the monthly price ²⁹.
3. Private or offline work. On-device dictation sends no audio to a provider, and its cost is paid in memory and battery rather than per minute. Our guide to offline speech to text on Windows explains which Windows tools keep working without a connection.
4. Cleanup on a budget. If you want an AI rewrite, a small model keeps that step under a cent an hour ³. A large model costs about 28 cents an hour ²³, close to some of the transcription rates above, so ask whether you need its extra capability for a dictation pass at all.

For a question-by-question view of the same trade-offs, see whether on-device dictation is cheaper than cloud. Prices change, so check each provider's current page before you decide, and use the dates on this page as the point of comparison.

How we calculated these numbers

Every calculated figure on this page is listed in its own fact with the method shown. The main assumptions are these.

• Speech rate. We use 150 words per minute, the working figure Brysbaert uses for speech ²².
• Tokens. We assume 1.3 tokens per English word. This is our estimate, not a sourced figure.
• Usage. The heavy user dictates two hours on each of 21 workdays, which is 2,520 minutes a month ⁴.
• Prices. Every price is a list price, read on October 16 and 17, 2026, and excludes promotions, volume discounts and taxes.

Where a provider's price depends on tier or product, we quote the tier that applies to a single user and name the product, so that you can recalculate with your own minutes.