1. Know what a local app still downloads and sends
In a local app the speech model is a file on your disk and recognition runs on your processor. That file has to arrive somehow, so almost every local app downloads a model on first run or from a model picker, often several hundred megabytes. After that, the dictation itself should need no connection.
Other traffic is normal as long as it is disclosed: update checks, license or trial activation, crash reports, and any optional cloud mode the same app offers. Many apps mix modes. Spokenly, for example, runs local Whisper and Parakeet models for free, also supports cloud engines through your own API keys or its Pro plan, and documents a local-only mode that blocks every network request. For any app, the question is which mode is active when you press the key.
- Once: the installer and the model file.
- Periodically: update checks and license check-ins.
- Per dictation: nothing, in a local mode.
- Only if you turn it on: cloud recognition, AI clean-up, or sync.
2. Size the model to your hardware and language
Two model families dominate local dictation. NVIDIA’s Parakeet TDT 0.6B v3 has 600 million parameters, transcribes 25 European languages with automatic language detection, outputs punctuation and capitalization, and is released under CC BY 4.0. OpenAI’s Whisper is a multilingual family whose reference table runs from tiny (39 million parameters, about 1 GB of GPU memory in OpenAI’s implementation) through base, small, and medium to large (1,550 million, about 10 GB), plus turbo (809 million, about 6 GB), an optimized large-v3 that OpenAI says is faster with minimal accuracy loss.
Desktop apps usually ship quantized builds that are smaller than those reference figures. Handy’s download list shows the practical range: Whisper Small at 487 MB, a quantized Whisper Large at 1,100 MB, Whisper Turbo at 1,600 MB, and Parakeet Unified EN 0.6B at 731 MB.
For live dictation, a bigger model is not automatically better: one that is slow on your machine adds a pause after every sentence. Start with the app’s default, dictate a realistic paragraph with your own names and numbers, and move up a size only if it fails. If your language is outside Parakeet v3’s list, choose a multilingual Whisper model and confirm the app offers it.
3. Check a Mac app with LuLu or Little Snitch
LuLu, from Objective-See, is a free, open-source firewall for macOS 10.15 and later that alerts you whenever a process makes a new outgoing connection, shows the process and its destination, and saves your allow or block decision as a rule. Little Snitch, from Objective Development, is a network monitor and application firewall that shows where your Mac connects, with connection alerts, connections grouped by app, domain, or country, and a traffic history.
Use either the same way. Let the dictation app download its model with alerts on. Then block its outgoing connections, quit and reopen it, and dictate several sentences: a local app keeps typing, while one that depends on a server fails or waits. Finally, review which hosts it tried to reach and when. A request at launch is usually an update or license check; a connection that coincides with every dictation means audio or text is leaving.
- Allow the first-run model download, then block.
- Restart the app after blocking so an open session does not hide a dependency.
- Line up connection times against the moments you dictated.
4. Check a Windows or Linux app from the command line
On Windows, find the app’s process ID in Task Manager’s Details tab, then run Get-NetTCPConnection -OwningProcess followed by that ID in PowerShell. Microsoft documents the cmdlet as listing TCP connections with their local and remote addresses, ports, and state, and -State Established narrows the output to live connections. Run it while the app is idle, again during a dictation, and again after, and compare the remote addresses.
On Linux, ss from iproute2 dumps socket statistics; its manual documents -t for TCP sockets, -u for UDP sockets, and -p to show the process using each socket. Run ss -tup while the app is idle and again while you speak, and look for new remote addresses owned by the dictation process. Then take the network down entirely, start the app, and dictate: the runtime result is the evidence that counts.
5. Cross-platform local apps and what to read before trusting one
If you use more than one operating system, one app on all of them keeps the hotkey, vocabulary, and habits the same. Handy is free and MIT-licensed, runs on Windows, macOS, and Linux, transcribes locally with Whisper or Parakeet, and documents a manual model install for networks that block downloads. Spokenly runs on Mac, Windows, Linux, and iPhone with free local Whisper and Parakeet models, optional cloud engines, and the local-only mode described above.
Before dictating anything sensitive, read three things in the app you choose: which model is active by default, whether any formatting, clean-up, or rewriting step calls a cloud model, and what the license or account system sends home. Then run the check for your operating system above, because a settings screen tells you what should happen and a connection log tells you what did.
6. IraVoice: what runs locally and what connects
IraVoice 1.2.2 runs on macOS 13+ on Apple silicon, Windows 10 or 11 (64-bit), and Linux (beta). Recognition runs on the computer with NVIDIA Parakeet; the default Parakeet Unified EN model is about 700 MB and downloads once on first run, and Parakeet v3 (25 European languages) and Whisper models are optional offline alternatives. On IraVoice Local, audio and text never leave the computer.
The connections it does make are the ones you should expect to see in a monitor: the one-time model download, plus trial and license check-ins and update checks to api.iravoice.com, which carry a salted hash of a hardware ID rather than the raw ID. There are no accounts and no product telemetry. The license checks in weekly and keeps working for 30 days without a connection, so an offline test lasts well beyond one session.
IraVoice Local costs $79 once for one computer, $109 for two, or $149 for three, after a 3-day free trial that starts with the first dictation and needs no email, card, or account. A separate IraVoice Cloud plan is listed as coming soon and is not on sale.
- Hold Right Command on a Mac, Right Ctrl on Windows, or Ctrl+Space on Linux.
- Corrections you teach it are kept as word pairs on the computer.
- Code names in developer tools come from the local Git project on Mac and Windows, never from source text.