On-device transcription vs cloud: privacy, cost and speed
What on-device transcription means technically, what cloud services do with your audio, the per-minute billing math, and the cases where cloud still wins.
On-device transcription means the speech model runs on your own Mac or iPhone and the audio never leaves it; cloud transcription means the audio is uploaded to a server, processed there, and the text is sent back. The choice affects who can hear your recordings, what you pay per hour of audio, and whether it works on a plane. This article explains the mechanics of both, does the billing math honestly, and names the situations where cloud is still the better tool.
What "on-device" means technically
A speech recognition system has two parts: the model, which is a large file of learned weights, and the runtime, which feeds audio through those weights to produce text. In a cloud service both live in a data centre. In an on-device app both live on your hardware.
Concretely, an on-device transcription app:
- Downloads the model weights once, usually at first launch. This is the one time it needs the network for the core job.
- Stores them on disk, typically a few hundred megabytes to a couple of gigabytes per model.
- Runs inference on the local CPU, GPU or, on Apple hardware, the Neural Engine. Frameworks like Core ML and MLX exist to make this fast on Apple Silicon.
- Writes the text to local storage. No audio, and no text, is serialised to a socket.
The practical test is simple. Put the device in airplane mode, or turn Wi-Fi off, and transcribe something. If it works, the recognition is on-device. If the app shows a spinner or an error, it is not, whatever the marketing says. On a Mac you can go further with a network monitor and watch for outbound connections while transcribing.
Two things that are sometimes confused with on-device processing:
- "Encrypted in transit" means the upload is protected from eavesdroppers on the way to the server. The server still decrypts and hears everything.
- "We do not store your audio" describes retention, not location. The audio was still processed on a machine you do not control, by a company that had it, however briefly.
Apple's own dictation is a useful reference point: for some languages it runs on the device, and for others it still sends audio to Apple's servers. Same feature, two very different privacy stories depending on the language you speak, which shows why the question is worth asking of every tool.
What cloud transcription services do with your audio
A cloud service has to do the following, and each step is a place where a policy applies:
- Receive the audio. It now exists on their infrastructure, and usually on a cloud provider's infrastructure beneath that, sometimes in another country.
- Process it. Larger providers may route work to subprocessors, which are listed, if at all, deep in a data processing agreement.
- Return the text, and often keep both audio and text for a period: to let you re-download, to debug, or to "improve the service". Training on customer data is usually opt-out or off for paid tiers, but the default varies by product and changes over time.
- Delete, eventually. Retention windows in consumer apps range from immediate to indefinite.
None of this is sinister. It is the normal shape of a cloud product, and reputable services document it. But it means that "is my transcript private" becomes a question about a vendor's policy, its subprocessors' policies, its breach history, and jurisdiction. With on-device processing that entire chain of questions disappears, because there is no vendor in the chain. For professionals who face confidentiality reviews, that is the whole point; private medical and legal dictation goes into the specifics.
The billing math
Cloud transcription is metered because every minute of audio costs the provider compute. That cost reaches you in one of two forms: a per-minute price on an API, or a monthly plan with a cap on minutes.
Raw APIs, as of writing, are cheap per minute: on the order of a fraction of a cent to a couple of cents, and what cloud transcription actually costs per hour compares the current providers side by side. Consumer apps built on top of them charge more and add a monthly minute allowance. To compare, work in hours of audio per month.
| Audio per month | At $0.01 per minute | At $0.02 per minute |
|---|---|---|
| 5 hours | $3 | $6 |
| 20 hours | $12 | $24 |
| 40 hours | $24 | $48 |
| 100 hours | $60 | $120 |
Someone transcribing two meetings a week is in the 5 to 10 hour band, where the cloud is genuinely inexpensive. Someone transcribing every call, lecture or interview is in the 40 to 100 hour band, where the meter matters, and where a capped consumer plan usually forces an upgrade to the next tier.
On-device transcription has no meter, because the compute is yours. You pay for software once, or by subscription, and then the hundredth hour costs the same as the first: nothing. For reference, ThinkScribe charges a flat $9.99 per month, $79.99 per year, or a one-time $199.99, for unlimited transcription across Mac, iPhone, iPad and Apple Watch. Against the table above, the flat price is roughly break-even at 8 to 16 hours a month and a clear saving past that. Below 5 hours a month, a cheap metered service can cost less, and it is fair to say so.
Latency, speed and working offline
Cloud transcription of a file means uploading it first. A one-hour recording at reasonable quality is tens of megabytes; on hotel Wi-Fi that upload can take longer than the transcription. Live cloud captioning adds a round trip on every phrase, which is why cloud dictation feels a beat behind your voice.
On-device transcription starts immediately and runs as fast as the local chip allows. On recent Apple Silicon, modern models such as Parakeet transcribe far faster than real time, so an hour of audio finishes in minutes, with no upload step at all. Older or smaller models are slower but still usable. The Whisper vs Parakeet comparison covers which model to pick for speed versus language coverage.
And then there is simply having no network. Trains, planes, hospitals, courtrooms, basements, field sites, a client's office where you do not want to join the guest Wi-Fi. On-device works in all of them. Cloud does not, and a queue-and-upload-later feature is a workaround, not a fix.
Where cloud transcription still wins
Being honest about this is more useful than pretending otherwise.
- Weak hardware. On a low-end laptop, an Intel Mac, or a phone with little memory, a large on-device model may not run at all, or runs slowly enough to be annoying. A data centre GPU does not care what you own.
- Huge batches. Transcribing a thousand hours of archive audio in one go is a parallel-servers job. Your Mac processes one file at a time.
- Shared workspaces. If a team needs the same transcripts, comments and search in one web place, a cloud product provides that by design. An on-device app can sync between your own devices, but it is not a collaboration server.
- Some languages and accents. The largest cloud models, trained on more data than anything that fits on a phone, are still ahead for low-resource languages and for very heavy accents. For English and the major European and Asian languages the gap has closed to the point where most users cannot tell.
- Platforms. If your work happens on Windows or Android, the on-device options are thinner and a browser-based cloud tool may be the only practical choice.
- Live captions for large events. Streaming captions to hundreds of viewers belongs on a server.
If you are in one of these groups, use the cloud with eyes open: read the retention policy, turn off training where offered, and do not upload anything you would not email.
How to tell if an app is really on-device
A short checklist, applicable to any transcription app:
- Does it work in airplane mode after setup? This is the test that matters.
- Does it require an account? A local app has no reason to know who you are.
- What does the App Store privacy label list? "Data not collected" is the honest answer for a local app, with an exception for optional crash reports.
- Does it name the models it runs? Open-source model names (Whisper, Parakeet, Qwen) are a good sign; a vague "our AI" can mean anything.
- Is there a per-minute limit? A meter almost always means a server.
The one-step private option
It lets you choose the engine, Apple's on-device models, Whisper or Parakeet, so the trade-offs above are yours to set. If you want to build the same thing by hand first, transcribe audio to text on a Mac walks through the open-source route.
Frequently asked questions
Is on-device transcription as accurate as cloud transcription?
For English and the major world languages, current on-device models such as Whisper and Parakeet are close enough to cloud services that the difference is rarely noticeable in everyday recordings. Cloud models retain an edge for low-resource languages and very heavy accents, and for very noisy audio the biggest models still do better.
Does offline transcription mean my audio is private?
If the app truly runs the model locally, the audio never leaves the device, so there is no vendor that could read, retain or leak it. The test is whether transcription works with the network off. Privacy then comes down to your own device's security and backups, which you control.
How much does cloud transcription cost per hour?
Raw API pricing as of writing works out at roughly $0.40 to $1.50 per hour of audio, and consumer apps typically bundle a monthly allowance of minutes at a higher effective rate. At 20 hours a month that is around $12 to $30. A flat-price on-device app costs the same regardless of volume.
Can on-device transcription work without internet?
Yes, that is the defining property. After the one-time model download, recognition, and in a fully local app the summaries too, run with no connection at all: in airplane mode, in a basement, or in a building where you will not join the Wi-Fi.
