Hold the dictation key, speak, release. The audio goes to a FluidAudio Parakeet speech-to-text model on your Mac's Neural Engine, the transcript gets deterministic punctuation and spacing cleanup, and the text is pasted at your cursor. There is no second model, no further inference call, and no network request anywhere in that path. It measured 127 to 173 ms from key release to text in our testing.
That is the entire operation, which is what lets this page make a narrow claim instead of a sweeping one. Most dictation is the other shape: audio leaves for a server, including Apple's own unless on-device processing is turned on in your Mac's dictation settings. You do not have to take our word for which shape ours is. Cut the network and watch dictation keep working.
Why most dictation apps use the cloud
Cloud STT models can be larger and trained on more data, and vendors tune them for edge cases like heavy accents, background noise, and domain-specific jargon. Running a large model in a data center is cheaper per inference because the cost is amortized across millions of users.
For the companies building these products, cloud processing has a second advantage: data. Audio recordings, transcription corrections, and usage patterns are valuable for model improvement. Training policies on these products have moved under public pressure before, and we covered one such episode in the Wispr Flow privacy incident. The structural point outlives whatever any current policy says: if the audio reaches the cloud at all, your training preference is a setting on somebody else's server rather than a property of the system.
The business incentive is clear. Cloud processing enables subscription pricing (ongoing server costs justify ongoing charges), provides training data, and allows using the largest possible models. The user's incentive is different - they want their words transcribed accurately and privately. These incentives aren't aligned.
What changed: the Neural Engine
Every Mac with Apple Silicon has a Neural Engine - dedicated hardware designed specifically for machine learning inference. Not a GPU. Not a CPU core repurposed for AI. A separate processing unit optimized for the matrix operations neural networks require.
Most of the time, the Neural Engine sits idle. Your CPU handles your apps, your GPU handles graphics, and the Neural Engine waits. Speech-to-text is a near-perfect workload for it - relatively small models, real-time audio processing, latency-sensitive results.
FluidAudio's Parakeet is a 0.6B parameter STT family compiled for the Neural Engine. The default preset is Parakeet TDT V2 for English; Parakeet TDT V3, covering 25 languages plus Japanese, is selectable in settings. With the model loaded as a keep-warm backend, end-to-end latency from key release to text at the cursor measured 127 to 173 ms in our testing. A cloud service has to finish uploading the audio before any of its own work begins.
Local isn't just more private. It's faster.
What private dictation actually means
"Private" gets overused in tech marketing. Here's what it means concretely for voice dictation.
When you speak into ToolPiper, the audio is captured by your Mac's microphone, processed by the Parakeet model on the Neural Engine, and the resulting text is inserted at your cursor. At no point does audio data leave your machine. The dictation path issues no network request, contacts no server, and has no cloud fallback. If your Wi-Fi is off, dictation works identically.
Some dictation apps use text from your active window to provide "context-aware" formatting in the cloud. That means whatever sensitive document, codebase, or communication is visible can travel to external servers alongside your audio. ToolPiper doesn't capture your screen and sends no image anywhere. It does read a little text immediately around your cursor, up to 256 characters either side, purely to punctuate and space the transcript correctly: whether the first word starts a sentence, whether a space is needed. That read happens on your Mac, password and secure fields are excluded from it, and the text is never stored, logged or transmitted. When it cannot be read, the transcript is simply cleaned up on its own.
Your dictation itself is not instrumented: no audio, no transcript, and none of the text you dictate is recorded, counted or sent anywhere. Being precise about the app as a whole: ToolPiper checks for updates, sends a daily activity count (on by default, with a stop switch in Settings under Activity) and optional usage counts you can switch off under Analytics. Signed out, the activity count carries no account or device identifier; signed in, ToolPiper also checks your license about every six hours, and that check is tied to your account. None of these carry prompts, transcripts, file contents or history. You still don't create an account, provide an email, or agree to terms that grant data processing rights. You download a DMG, install it, and use it.
Who needs this
Some people want privacy as a matter of principle. For many others, it's a professional requirement.
Attorneys dictating case notes, contract terms, or legal strategy into a cloud-based transcription service create a data exposure most compliance teams would flag immediately. Healthcare workers face similar constraints: HIPAA requires a BAA with any service that processes protected health information, and most dictation apps don't offer one. Transcribing on your own machine removes that third party from the dictation step. It is not by itself a compliance program, and none of this is legal advice about your obligations. Developers dictating about proprietary software may be leaking trade secrets, unreleased feature names, or security-sensitive details through a third party's infrastructure.
Financial professionals, journalists protecting sources, researchers handling unpublished findings - the list goes on. But you don't need to be in a regulated industry. Most people, if asked directly, "would you like your dictation app to send what's on your screen to a server alongside your audio?" would say no. The default should be privacy. Cloud processing should be the opt-in.
The bigger picture: AI shouldn't live in someone else's data center
Private dictation isn't a niche concern. It's the leading edge of a larger question about where your AI runs.
The current model - upload your data, pay monthly, trust the provider - exists because local hardware couldn't match cloud inference quality. That gap is closing fast. STT proves it. Voice input is latency-critical. Sending audio over the wire adds hundreds of milliseconds before processing even starts. On-device STT on the Neural Engine measured 127 to 173 ms from key release to text in our testing. The local version isn't just more private - it's physically faster because there's no network to traverse.
Think about what a cloud STT subscription, $12 a month billed annually or $15 month-to-month at the time of this writing, actually buys you. Someone else's GPU time to run a model that fits on hardware you already own. Your Mac already has the Neural Engine to run it. You paid for that silicon at the Apple Store. Paying again, monthly, for a cloud service to do the same work slower while capturing your audio - that's a transitional pricing model, not an enduring one.
Local models get smaller and faster every quarter, and each generation is more accurate than the last. In a few years, paying a subscription for speech-to-text will seem as dated as paying per minute for long-distance calls. The capability was always going to move to the device.
The accuracy question
The strongest argument for cloud dictation is accuracy in hard conditions. Cloud vendors train large models on more data, run them with more compute per request, and put real work into heavy accents, noisy rooms and rare vocabulary. Wispr, for example, reports that its Canto model cuts error rates in noise, wind and music to between 5 and 10%.
The Parakeet model on the Neural Engine delivers Whisper-class accuracy for standard English in reasonably quiet environments. For the majority of use cases - writing emails, taking notes, drafting messages, commenting code - local accuracy is indistinguishable from cloud.
The question is whether a cloud vendor's work on the hardest audio is worth sending all your audio - including the 90% that local models handle perfectly - to external servers. For most users, it isn't.
How ToolPiper handles private dictation
ToolPiper is a macOS app with two voice input modes, both running on-device. Each is on its own key, and every voice key is rebindable in Settings under Hotkeys, to Command, Option, Control or Shift, on the left side, the right side or either, or switched off entirely. The shipped defaults stay on the right side, and that is what the instructions below use.
Push-to-talk dictation (Right Option by default) - hold the key, speak naturally, release. Parakeet transcribes on the Neural Engine and inserts text at your cursor in whatever app is focused. 127-173 ms end-to-end. Works offline.
Push-to-command (Right Command by default) - hold the key, speak a natural language instruction, release. The STT engine transcribes, a local LLM interprets the command against 26 action domains, and your Mac executes it. "Turn on dark mode." "Set volume to fifty percent." "Snap this window to the left half." All local.
Both modes are free in ToolPiper, with no account. The STT engine and local LLM underneath are part of the free runner: transcription, chat, and the push-to-talk key in every app all cost nothing.
Data flow comparison
| App | Processing | Audio sent to cloud | Screenshots captured | Account required | Price |
|---|---|---|---|---|---|
| ToolPiper | On-device (Neural Engine) | No | No | No | Free |
| Wispr Flow | Cloud (their servers) | Yes, always | Not stated; uses active-window text with Context Awareness on | Yes | Free tier; Pro $12/mo annual, $15/mo monthly |
| Apple Dictation (default) | Cloud (Apple) | Yes, by default | No | Apple ID | Free |
| Apple Dictation (on-device) | On-device | No | No | Apple ID | Free |
| Otter.ai | Cloud | Yes | No | Yes | $10-25/month |
| Whisper.cpp | On-device | No | No | No | Free (CLI only) |
Setup
Download ToolPiper (DMG, direct download). Install ToolPiper from modelpiper.com to get the STT engine. Grant Accessibility permission when macOS prompts you. Hold Right Option, or whichever key you bound dictation to, speak, release. Text appears at your cursor. Everything ran on your Mac.
ToolPiper is part of the ModelPiper family of local AI tools for Mac. See also: Wispr Flow Alternative, Push-to-Talk AI on Mac, Voice Coding on Mac.
