What on-device speech recognition means

On-device speech recognition means your device converts spoken audio to text using its own processor, without sending the audio to a server. Parlo works this way: Apple's SpeechAnalyzer and SpeechTranscriber transcribe on your iPhone or Mac, then an on-device language model cleans up the text. Apple's own dictation feature can work either way, depending on a setting.

By Eric Friedman. Last updated 2026-09-08.

On-device speech recognition is exactly what it sounds like. The device in your hand or on your desk does the work of turning speech into text, rather than sending your audio somewhere else to be processed.

What does on-device speech recognition mean?

Your phone or computer has enough processing power to run a speech-to-text model itself. When an app uses on-device speech recognition, your spoken audio is converted to text by that local model, using your device's own processor, without a network request carrying the audio anywhere. The text that comes out can still be used however the app wants afterward, including being sent somewhere later, but the conversion from speech to text itself does not require that step.

How does Parlo do it?

Parlo transcribes with Apple's SpeechAnalyzer and SpeechTranscriber, the on-device speech engine built into iOS and macOS. It is the only transcription engine Parlo uses: there is no cloud engine, and no option to bring your own API key for a different one. The same on-device engine family sits behind the newest version of Apple's own system dictation. You can try it during the beta.

What runs on-device besides transcription?

Cleanup. After transcription, an on-device language model removes filler sounds, false starts, and repeated words. Spoken formatting commands like "new line," "new paragraph," and "parentheses ... end parentheses" apply even with this cleanup step off; "quote" and "end quote" need the model turned on. A length guard checks the cleaned result against what you actually said, so an implausibly long rewrite falls back to a simpler pass instead. With cleanup off, Parlo still strips filler sounds and applies the deterministic commands.

Does on-device mean the app never needs the internet?

Not entirely, but close. The first time a device runs on-device speech recognition, it may need to download the model itself, which needs a one-time connection. After that download, transcription runs locally without a similar need. Separately, beta builds of Parlo refresh their text-transform prompts once a day from parlokey.com, a fetch that carries no user data and has nothing to do with the audio or transcript from a dictation.

How is this different from cloud speech recognition?

A cloud speech recognition system sends your recorded audio to a server, converts it to text there, and sends the text back. That trip introduces a step where the audio exists somewhere other than your device, even briefly. On-device recognition skips that trip entirely for the speech-to-text conversion itself. What happens to the resulting text afterward, such as an optional encrypted sync, is a separate question from where the speech-to-text conversion ran.

Does Apple's own dictation work the same way?

Not always. Apple's privacy page for Ask Siri and Dictation states that "your device will indicate in Keyboard Settings if your audio and transcripts are processed on your device and not sent to Apple servers." That means Apple's dictation can run either on-device or through Apple's servers, depending on your device, language, and that setting. Parlo's pipeline does not have that second path: transcription and cleanup run on-device every time, on iPhone and on Mac.

Why does on-device processing matter beyond privacy?

Skipping the trip to a server also means transcription does not depend on your connection quality once the model is on the device. It has nothing to do with accuracy by itself: Parlo does not publish an accuracy number, and neither does Apple for its own dictation. The honest comparison is to try dictation on your own voice and vocabulary rather than trust a claimed percentage from either side.

Does on-device processing affect what happens if you delete a note?

Yes, in a small but real way. Because Parlo never uploads your audio, there is no recording sitting on a server that also needs to be deleted when you remove a note from the app. Deleting a note moves it to a device-local trash for 30 days, and that trash is never synced, so it stays on the one device where you deleted it. If history sync was on for that note before you deleted it, the synced copy is handled through Settings, where you can delete synced dictations, your dictionary, and preferences from iCloud at any time.

Does on-device speech recognition mean there is no server involved anywhere?

Not quite. Transcription and cleanup, the steps that turn your voice into cleaned-up text, run entirely on-device. Two separate things can still touch a network: a one-time model download on first run, and a daily fetch of text-transform prompts in beta builds, which carries no user data. If you turn on history sync, the encrypted transcript later travels to your private iCloud database. By that point it has already been encrypted on your device, so what leaves is ciphertext, not your recognizable words.

How do you know a claim like this is accurate for a specific app?

Look for the app naming the exact engine it uses and where each step runs, the way this guide names Apple's SpeechAnalyzer and SpeechTranscriber for transcription and Apple's FoundationModels for cleanup. A vague claim like "your data is safe" without naming a specific step is harder to verify. A claim naming which part runs locally, which part is optional, and which part, if any, involves a network request is easier to check. Apple's own privacy pages follow this same pattern for Dictation, naming the specific setting that controls which path your device uses.