In short
- With cloud transcription your audio is uploaded to a server; with on-device transcription it never leaves your computer.
- What happens to uploaded audio depends on the provider's terms: some never keep it, some can use it to improve their models unless you opt out.
- Open models like OpenAI's Whisper and Apple's Speech framework make good transcription possible on a Mac.
- On-device costs a one-time model download and some processing time; cloud costs privacy and a dependency on someone else's terms.
On-device transcription turns your speech into text on your own computer: the model is downloaded once and your audio never leaves the machine. Cloud transcription uploads your audio to a server, which sends the text back. The difference that matters is what happens to your voice after the upload, and that depends entirely on the provider’s terms.
Below: how both work, what three large cloud providers say they do with your audio, and what you trade with each.
Why it matters for client videos
Subtitles make a video easier to follow on mute and easier to skim. To make them, something has to listen to everything you said: the client’s name, the budget you explained, the bug in their checkout.
With a cloud service, that audio goes to the app you use and often to a transcription provider behind it. With on-device transcription, it stays on your Mac. For a public tutorial the difference may not matter. For a video about a client’s project, under an NDA or under GDPR, it can.
How cloud transcription works
The steps are the same everywhere:
- Your app uploads the audio (or the whole video) to a server.
- A speech recognition model on that server converts it to text.
- The text comes back, and the app attaches it to your video.
What happens to the audio after step 3 is where providers differ. Here is what three large providers say in their own documentation:
| Provider | Used to train models? | Kept afterwards? |
|---|---|---|
| OpenAI API, audio transcriptions | ✗ Not unless you opt in to share data | ✗ No abuse monitoring or application state retention for this endpoint |
| Google Cloud Speech-to-Text | ✗ Only if you opt in to its data logging program | ✗ By default it doesn’t log customer audio or transcripts |
| Amazon Transcribe (AWS) | ~ AWS AI services may use content for service improvement, including model training, unless you opt out | ~ May be stored, possibly in another AWS Region, unless you opt out |
From each provider's own documentation, checked November 2026. Settings and terms can change.
Two things follow from this. First, “cloud” isn’t one thing: the same audio can be handled very differently depending on who processes it. Second, you usually don’t deal with these providers directly. The recording app you use picks one, and its own terms and subprocessor list tell you which. Loom, for example, says its transcription draws on OpenAI’s GPT and Google’s Gemini models, a self-hosted Whisper and the transcription provider Speechmatics, which its subprocessor list shows with a US location. We go through the details in Loom and GDPR.
How on-device transcription works
On-device transcription runs the same kind of model on your computer instead:
- The app downloads a speech recognition model once.
- When you record, the model runs on your Mac’s processor and turns the audio into text.
- Nothing is uploaded for transcription. It works offline.
Two families of on-device models are common on the Mac.
Whisper
Whisper is “a general-purpose speech recognition model” released by OpenAI, with its code and model weights under the MIT license. OpenAI describes how it was trained in a research paper published with the model. Because the weights are public, apps can ship it and run it locally. For a hands-on look, see running Whisper on a Mac.
It comes in several sizes, and the choice is a trade between speed and accuracy. The official repository lists:
| Model | Parameters | Memory needed (approx.) | Relative speed |
|---|---|---|---|
| tiny | 39 M | 1 GB | about 10x |
| base | 74 M | 1 GB | about 7x |
| small | 244 M | 2 GB | about 4x |
| medium | 769 M | 5 GB | about 2x |
| large | 1,550 M | 10 GB | 1x |
| turbo | 809 M | 6 GB | about 8x |
From the official openai/whisper repository, checked November 2026. Speed is relative to large.
The repository notes that turbo is a faster version of large-v3 with only a small loss in accuracy, and that English-only versions of the smaller models generally do better on English.
Apple’s Speech framework
Apple’s Speech framework has offered on-device recognition for years. In the older SFSpeechRecognizer API, an app can set requiresOnDeviceRecognition to prevent a request “from sending audio over the network”, with the caveat, in Apple’s words, that “on-device requests won’t be as accurate”. If that flag isn’t set, some languages may need an internet connection.
With macOS 26, Apple added SpeechAnalyzer and SpeechTranscriber. Their models are downloaded from Apple’s servers once, managed by the system and shared between apps, so a second app that needs the same language doesn’t download it again.
The trade-offs
| On-device | Cloud | |
|---|---|---|
| Your voice stays on your computer | ✓ | ✗ Uploaded to the app, often to a provider too |
| Free of someone else’s terms | ✓ | ✗ Their terms apply, and they can change |
| Works offline | ✓ After the first download | ✗ |
| Ready on first use | ✗ One-time model download; larger models mean larger downloads | ✓ Nothing to download |
| Speed | Depends on your Mac and the model size | Depends on upload time and server load |
| Accuracy | ~ Good for clear speech; smaller models make more mistakes | Large models can do better with noise, accents and jargon |
| Battery and heat | ~ Uses your processor while it runs | ✓ Light on your device, but needs an upload |
For most client videos (one voice, a decent microphone, a quiet room) the accuracy gap is small. Where cloud models still have an edge is hard audio: several people talking over each other, background noise, heavy jargon. Whichever you use, a better microphone helps every model. Our guide to adding subtitles to a screen recording on Mac covers the practical side.
What it means under GDPR
A recording of your voice, and the transcript of what you said, are personal data whenever they relate to an identifiable person. Uploading them to a transcription provider makes that provider a processor, which means a contract, a place in your records and, if it’s outside the EU, a transfer mechanism. On-device transcription doesn’t remove your other obligations, but it takes one processor out of the chain. We go through the roles in GDPR and client videos, and our GDPR-friendly screen recorder page shows how JustRec handles each step.
Which one to choose
Choose on-device when the audio is about clients, contracts or anything under NDA, when your clients are in the EU and ask where their data goes, or when you often work offline.
Choose cloud when the content is public anyway (a tutorial for your website), when the audio is hard and accuracy matters more than anything else, or when you need features that run on the transcript in the cloud, like AI summaries.
Many people will end up using both. The point is to know which one a given app uses, before you record the client’s budget into it.
Subtitles are also what gets a video watched on a phone in a quiet office: more on that in subtitles get your videos watched on mute.
Sources
- OpenAI: Whisper repository, checked November 2026.
- Radford et al., the Whisper paper (arXiv 2212.04356), checked November 2026.
- Apple Developer: requiresOnDeviceRecognition, checked November 2026.
- Apple Developer: SFSpeechRecognizer, checked November 2026.
- Apple Developer: SpeechAnalyzer and AssetInventory, checked November 2026.
- Atlassian Support: Loom AI features and Atlassian Sub-processors, checked November 2026.
- OpenAI API: Data controls in the OpenAI platform, checked November 2026.
- Google Cloud: Speech-to-Text data logging, checked November 2026.
- AWS Organizations: AI services opt-out policies and list of supported services, checked November 2026.
Frequently asked questions
What is on-device transcription?
Speech-to-text that runs on your own computer or phone. The recognition model is downloaded once, and your audio is turned into text locally without being sent to a server.
Is on-device transcription as accurate as cloud transcription?
Often close, but not always. Large cloud models can do better with noisy audio, strong accents or rare terms, and Apple's own documentation notes that on-device requests in its older speech API won't be as accurate. For a clear voice in a quiet room, current on-device models do well.
Do cloud transcription services keep my audio?
It depends on the provider and its settings. OpenAI says audio sent to its transcription API isn't used for training or kept for abuse monitoring, Google Cloud Speech-to-Text doesn't log audio unless you opt in, and AWS says its AI services may store content for service improvement unless you opt out.
Can Whisper run offline on a Mac?
Yes. Whisper is an open-source speech recognition model from OpenAI, released under the MIT license, and once the model is downloaded it can run locally without an internet connection.
Does on-device transcription use a lot of battery?
It uses the processor while it runs, so a long transcription on battery will drain it faster than idle work. Smaller models are faster and lighter; larger ones are more accurate but need more memory and time.
