On-device transcription is more private because your audio is never uploaded. There’s no copy on someone else’s server to be breached, handed over in response to a legal request, or used to improve a model, and no retention policy you have to take on trust. Cloud transcription can be well secured, with encryption, audited controls and no-training commitments, but the provider still holds your recording under its rules. Cloud services still lead on collaboration, speaker labels and some very hard audio. For voice notes, journals, and any meeting you wouldn’t forward to a stranger, on-device should be the default.
What happens to your audio in each approach? #
With on-device transcription, the speech model runs on your phone or computer. The microphone records to local storage, the model turns it into text on your own processor, and the result stays in the app until you share or delete it. It works with the network switched off.
With cloud transcription, the app uploads the recording (or streams it live) to the provider’s servers. The provider transcribes it there and sends the text back, usually keeping the audio, the transcript or both in your account so you can search and share them later.
| Factor | On-device | Cloud |
|---|---|---|
| Does the audio leave your device? | No | Yes |
| Works without internet | Yes, after the model download | No |
| Copy held by a third party | None | Yes, under its retention policy |
| Exposure to a provider breach or legal request | None for the provider to hand over | Possible |
| Used to train models | No | Depends on the provider and your settings |
| Speaker labels, summaries, team sharing | Limited | Usually strong |
| Typical pricing | Free or one-time | Monthly subscription or per minute |
What can a cloud transcription service do with your recording? #
It depends entirely on the company, which is the core of the problem: you’re relying on policy rather than architecture. Things to look for in any provider’s terms:
- Storage and retention. Most keep your files until you delete them, because a searchable archive is the product. Some let you delete at any time. Sonix, for example, says deleted files have both audio and transcripts wiped from its platform.
- Model training. Policies vary. Sonix states it doesn’t use customer data for training. Wispr Flow lets you choose whether your data can be used to improve its models, and whether it stores audio and transcripts at all. Others use de-identified data by default unless you opt out.
- Subprocessors. Many services send audio to a separate AI provider or cloud host, which adds another company to the chain.
- Legal requests. Data a company holds can be subject to subpoenas and court orders in the countries where it operates.
None of this means cloud providers are careless. The better ones publish security audits and let you delete everything. But a well-guarded copy is still a copy, and it exists as soon as you press upload.
What on-device transcription doesn’t protect you from #
Keeping audio on your phone moves the responsibility to you. Be clear about the gaps:
- A lost or unlocked phone. Use a strong passcode and biometric lock. Device encryption only protects you if the phone is locked.
- Backups and sharing. Once you share a transcript by email or save a backup to a cloud drive, that copy lives wherever you put it.
- The app’s other connections. Check what else the app does online. Private Transcribe, for instance, only touches the network for the one-time model download, for ads in the free version, and for Apple or Google purchase validation. None of that involves your recordings.
- Marketing that says “private” but means “encrypted.” Some apps call cloud processing with encryption “private.” It isn’t on-device.
When is cloud transcription the better choice? #
Be honest about the trade-off. Cloud services are still the better tool when you need:
- Shared meeting notes that a whole team edits, comments on and searches.
- Reliable speaker names pulled from the meeting platform.
- Integrations with calendars, CRMs and project tools.
- Human-verified transcripts for court filings or publication.
- Very large volumes, such as thousands of hours of archive audio, where your own hardware would take too long.
If none of your recordings are sensitive, the convenience may well be worth it. Most people have both kinds of audio, and it’s reasonable to use both kinds of tool.
Why voice notes deserve on-device transcription #
Voice notes are the most personal audio most people make: half-formed ideas, journal entries, symptoms to tell a doctor, a rant about a colleague. They’re also short and made on a phone, which is exactly where on-device transcription now works well.
Your options on a phone:
- iPhone keyboard dictation processes speech on the device in many languages. Settings > General > Keyboard shows whether yours does.
- Google’s Recorder app on Pixel phones transcribes entirely on the device, with no internet connection needed.
- Private Transcribe runs OpenAI’s Whisper model on any recent iPhone or Android phone. You record or import a voice memo, get the text within seconds, and nothing is uploaded. It needs no account, keeps a searchable history on the phone, and can delete each recording automatically once it’s transcribed, keeping only the text. Settings > Clear All Data wipes everything, and since there’s no server copy, that really is everything.
We cover the step-by-step routine for voice memos in private voice notes to text on iPhone. There’s more on automatic deletion in speech-to-text apps that delete audio after transcription.
Is on-device transcription less accurate? #
Less than it used to be. Whisper-based models running on a recent phone or laptop are competitive with cloud services on clear speech, and the largest on-device models handle accents and moderate noise well. Cloud services keep an edge in speaker separation and on very noisy, many-speaker recordings. For a detailed comparison, see private transcription accuracy vs online services.
How much does each cost? #
On-device tools tend to be free, open source or one-time purchases, because the provider isn’t paying for servers each time you transcribe. Cloud services charge for that compute. Otter, for example, lists its Pro plan at $16.99 a month, or $8.33 a month billed annually, with a free tier of 300 minutes a month capped at 30 minutes per conversation (Otter pricing). If you transcribe every day, a one-time purchase pays for itself quickly.
How to check whether an app transcribes on your device #
- Airplane mode test. Download any model the app needs, switch on airplane mode (Wi-Fi off too), and record. If you still get a transcript, the recognition is local.
- Account requirement. An app that insists on an account before it will transcribe usually needs one because your recordings sync to its servers.
- Privacy label. Check the App Store privacy details or the Google Play data safety section for audio data collected.
- Policy wording. Phrases like “processed on our servers,” “subprocessors” or “to improve our services” mean cloud processing.
For a list of apps that pass the airplane mode test on every platform, see speech-to-text apps that work completely offline.
Frequently asked questions #
Is on-device transcription as accurate as cloud transcription? #
For clear speech in major languages, a mid-size or large on-device model is close to cloud quality. Cloud services still do better at labelling speakers and with very noisy group recordings. Choosing a larger on-device model closes much of the gap.
Can I make a cloud transcription app private with the right settings? #
You can reduce the risk by turning off model training, deleting files after transcription and choosing a provider with audited security. But the audio is still uploaded and processed by a third party, so it isn’t private the way on-device processing is.
Do cloud transcription services train AI on my recordings? #
Some do, some don’t, and some let you choose. Sonix says it doesn’t use customer data for training, and Wispr Flow gives you a setting for it. Read the policy of any service you use, and look for an opt-out.
What happens to my transcripts if I lose my phone? #
With on-device apps, the transcripts are only on that phone unless you made a backup. A passcode keeps them locked, but you’d lose them along with the device. If you need them long-term, export the important ones to storage you trust.
Does on-device transcription work without internet? #
Yes. Once the speech model is downloaded, on-device apps transcribe in airplane mode, on a plane or anywhere without signal. That’s also the easiest way to prove an app really is on-device.