If you're a lawyer, a doctor, or an engineer working on proprietary code, you may be bound by rules about where client, patient or company information can go. Your voice counts: a dictated email contains everything the email does.
So the first question to ask about any dictation app is simple: where does my audio go? There are three answers: to a server in the cloud, nowhere (it stays on your device), or a mix of the two, which is what people mean by hybrid voice architecture.
What Happens to Your Voice in a Cloud Dictation App#
When you dictate with a cloud tool, your speech is recorded, sent over the internet to a server, transcribed by a speech model and returned as text. Many apps then pass that text to a language model for cleanup.
What happens next depends on the provider's policies: whether the audio is stored and for how long, whether people or models ever see it, and which other companies process it. Those are commitments, not technical guarantees, so read them, and check whether your organisation already approves the service.
Cloud processing has real advantages. Large models run quickly on server hardware without using your laptop's memory or battery, and the app behaves the same on any computer. The trade-offs: you need a connection, and you depend on how the provider handles your data.
Where LumeVoice fits: LumeVoice is a cloud app. Speech is transcribed by Whisper large-v3-turbo, run by Groq (with Mistral Voxtral as a fallback), and a language model cleans up the text. Audio isn't stored after processing, your transcription history stays on your device, and nothing you say is used to train a model. It has no offline mode.
What Apple Silicon Changed for Local Dictation#
Apple's M-series chips made local speech recognition practical on a laptop. They combine a Neural Engine for machine learning with a capable GPU and memory shared between them, which suits models like Whisper.
Open-source engines made use of it: whisper.cpp (MIT-licensed) and Argmax's WhisperKit run Whisper models directly on a Mac. Apps such as MacWhisper, Superwhisper (with local models), VoiceInk and Buzz transcribe on your own machine, and Apple's built-in dictation can process speech on your Mac (Keyboard settings show which languages it handles on the device).
When transcription happens locally, the audio stays on your Mac, and dictation keeps working with the Wi-Fi off. The trade-offs are practical rather than about privacy:
- Hardware. Bigger models need more memory. MacWhisper's own page says its Medium and Large Whisper models need a Mac with more than 8 GB of RAM to work well.
- Speed and battery. Running a large model takes power, and an older Mac may be noticeably slower than a server.
- Cleanup. Local apps vary in whether they remove filler words and apply corrections, and in whether that step runs locally too.
How Hybrid Architecture Works#
A hybrid design splits dictation into two layers:
- The transcription layer (on your device): you speak, and a local model turns the audio into raw text. The audio never leaves your computer.
- The reasoning layer (local or cloud): the raw text, not the audio, can optionally go to a more capable language model to remove filler words, fix formatting, restructure notes or translate.
You control the dial. For a brainstorm, you might switch on a cloud reasoning layer for better cleanup. For work that must never leave your machine, keep both layers local, or use a fully local tool.
Be clear about what "hybrid" doesn't mean: once text goes to a cloud model, that text has left your machine, even if the audio hasn't. A dictated client email is just as confidential as text as it was as audio.
| Design | Where your audio goes | Where your text goes | Works offline | Examples |
|---|---|---|---|---|
| Cloud | A server | A server | No | LumeVoice, Wispr Flow, Aqua Voice, Otter.ai |
| Local | Stays on your device | Stays on your device | Yes | MacWhisper, VoiceInk, Buzz, Apple Dictation (for the languages Keyboard settings list as on-device) |
| Hybrid | Your device, with a local model | A cloud model, if you choose one | Partly | Superwhisper (local or cloud models), Spokenly (free local models, cloud Pro plan), Amical (local-first, cloud fallback) |
Swipe to see all columns
Questions to Ask Before You Choose#
Whatever the architecture, these questions settle most privacy decisions:
- Where is my audio processed: on my device or on a server?
- Is the audio stored after processing, and if so, for how long?
- Is anything I say used to train models?
- Where does the text go for cleanup?
- Where is my history kept: on my device or in a cloud account?
- Does it work offline?
- Does my employer, client or regulator allow it? For regulated work (HIPAA, GDPR, legal privilege), that answer outranks everything else.
For LumeVoice, the answers are: on a server; no; no; to a cloud language model; on your device; no; and the last one is up to your organisation.
For work that legally can't leave your machine, the safe choice is a tool that transcribes on-device, such as MacWhisper or Superwhisper with local models, with any cleanup kept local too, because no setting removes the upload step from a cloud transcriber. Our guide to private dictation for legal and medical work goes further.
Cloud Dictation, Handled Carefully
LumeVoice is cloud-based, and upfront about it. Your audio is transcribed by Whisper large-v3-turbo and isn't stored after processing, your transcription history stays on your device, and nothing you say is used to train a model. In return you get a large Whisper model and automatic cleanup without loading your own machine, on Mac, Windows, iPhone and Android.
Further Reading#
- What agentic AI voice dictation is, and how it works
- Private dictation for legal and medical work
- Does Wispr Flow work offline?
- Best open-source Whisper apps for Mac
- The most accurate dictation app for Mac
- Best Bilingual Dictation Apps in 2026: Which Ones Switch Languages for You
- Best Speech to Text API in 2026: OpenAI, Deepgram, AssemblyAI, Google, Azure and AWS
- Speech to Text Guide: How It Works and How to Choose a Tool
- Voice Recognition Software: What It Is and Which Kind You Need
