The app is built to work the instant you press the key, with or without a network. Unlike cloud-based dictation, which routes every sentence through the internet, this one runs everything on your machine. Model weights load once at home or office, and after that your phone stays silent, your data stays private, and your dictations land with no latency waiting for a distant server to wake up.
What you need to know before your first offline dictation
The model downloads once, automatically. When you open Coii VoiceInput for the first time, or after an update, the model weights are fetched from a public host and stored on disk. This is the only large network transfer the app ever makes. After that, every dictation is purely local — nothing leaves the machine, and nothing needs to come back in.
Activation happens online, but does not have to repeat. The licence you buy is tied to the machine and the account that activated it. Activation requires a network connection the first time and the first time you move to a new machine, but after that, the app does not phone home between dictations. On an aeroplane, in a SCIF, or in a parking garage with no signal, dictation works exactly as it would in your home office.
A clock check runs once at launch. The app validates the time against a remote server to catch licence clocks that have been tampered with. If it cannot reach the internet, the launch still succeeds and dictation works — the check is deferred and tries again next time you have a network. This is not a permission check and not a licence check; it is a time check, and the failure mode is simply to let the offline session proceed.
After that, everything is local. While you hold the key, your mic input goes to the engine on the machine. When you let go, the transcript lands in the field you chose. Both steps happen on the machine. Your clipboard, your accessibility API, your field focus — all local. The only way an offline dictation differs from an online one is that it is instantaneous.
Where the architecture makes this possible
Most transcription apps are cloud-first. The audio from your microphone goes to a remote server, the model runs there, and the result comes back. That model is expensive to run and needs to live where the CPU and memory are, so even the ones that support offline use often do it by running a separate, smaller model locally — which is why offline mode is sometimes slower, less accurate, or missing features.
- 1Key downthe target is chosen here
- 2You speaklevel meter, over your work
- 3Key up
- 4Written to historybefore the engine is asked
- 5Engineon your Mac
- 6Typedabout half a second
This app runs the full model on your machine from the start. There is no smaller model for offline and a bigger one for cloud. There is no fallback to a worse engine when the internet is down. The machine is the server, and the boundary between online and offline is simply whether the clock check succeeded or the launch just moved on.
The difference matters most when you are committing words you might need later. Because the transcript is written to your history before the engine is asked anything — offline, online, or somewhere in between — a slow model, a crashed engine, or a network timeout cannot cost you the sentence. The local backup is already there.
Setting up for offline dictation
If you mostly work online: Install the app normally, let the model download at home or the office, and forget about it. You are set for offline from that point onward. When you leave for a trip or go somewhere with no service, the app will work exactly as it does at home.
If you work in a place with no internet on most days: The first time you use the app, do it somewhere with internet so the model downloads and installs. After that, you never need a network connection again — bring your Mac with the app already installed, turn on Airplane Mode, and dictate. Licence activation only happens once per machine, so if you have already activated elsewhere, a fresh install on a machine in a no-internet location will need activation before you can start; activate somewhere with a network (on a mobile hotspot, at a library, or on a trip to a connected place) and the machine is set for offline use after that.
If you are activating on a new machine with no network available: You will need a connection for the first launch. This is unavoidable — the licence is bound to the machine, and the binding requires verification. After activation, offline dictation is unrestricted.
What changes when you work offline
Very little. The speed and accuracy you get offline are the same as what you get online, because it is the same model and the same machine. The only difference is the absence of a round-trip to a server: your results come back in the same time it takes to transcribe, which is about half a second after you let go of the key.
Your microphone still closes instantly. Hold is the default; as soon as you let go of the key, the mic closes. Nothing is listening between dictations. The microphone closes the moment you release, whether you are online, offline, on a plane or in a bunker. This is not a setting; it is how the app works.
Your history still saves everything. The raw transcript and the processed transcript both write to the database before the engine is asked anything. If you work offline for an hour, the history row is complete the moment the dictation lands — your words are already backed up.
Filler removal still requires you to turn it on. Filler removal is off by default in all contexts. If you want um and uh removed, enable it in settings, and it works the same whether you are online or offline. If you do not enable it, they stay in the transcript.
Your dictionary works the same. The entries you added to your dictionary are on the machine. They work offline as well as online, before recognition and over the transcript — your custom terms and rewrites travel with your license to every machine you activate it on.
Offline scenarios this handles
Aeroplanes: Download the model before boarding. Activate licence on the machine beforehand. Plane WiFi, if available, is not needed — Airplane Mode works fine.
Remote locations without internet: Coffee shops, libraries, national parks, or work sites where the network is out or unreliable. The app works the same whether connectivity is intermittent or completely absent.
Secure facilities: SCIFs, clean rooms, and locked-down networks where transcription software cannot phone home. No uploads, no outbound connections, no data leaving the machine.
Privacy-sensitive dictation: Anything you would rather not send to a cloud service — legal documents, medical notes, trade secrets, or simply writing you want to keep to yourself.
Travelling with a laptop: Because the model is on the machine and the only network requirement is the initial download and activation, carry your Mac anywhere and dictate without thinking about where you are or whether you have signal.
Unreliable connectivity: If your internet cuts out, the app keeps working through the interruption. Start a dictation online, and if the connection drops between your release and the text landing, the sentence is already in your history anyway — the network failure is behind you, not ahead.
The licence and offline use
Every licence includes offline use. You are not buying "online dictation" and then paying for an "offline" tier — you buy one licence and get all of it, everywhere. The only restriction is that activation requires a network the first time. After that, the machine is tied to your account and knows to allow dictation whether it can reach the internet or not.
If you buy a three-device licence, each machine you activate is one slot. All three work offline after activation. No per-session cost, no metering, no cloud-based tracking of which machine is online at what time.
Offline vs. cloud-based dictation: The architectural difference
Most dictation apps, including Wispr Flow, are built on cloud infrastructure. Your audio goes to a remote server, the transcription model runs there, and the result comes back to you. This architecture means:
- Offline mode, if offered, runs a separate, often smaller model locally
- Cloud mode benefits from frequent updates and more compute
- Your audio leaves your machine on every dictation
- You depend on network availability for the full feature set
This app is built on a different principle: the full model runs on your machine from the start. There is no smaller offline model and a bigger cloud model. There is no fallback to worse accuracy when you are offline. The machine is the server, and offline is not a degraded mode — it is the only mode.
This means your dictations never leave the machine, your results are identical whether you are online or off, and the only network requirement is once-per-session at launch for a clock check. Read more about what makes this different or see the full comparison with the most common cloud option in the Wispr Flow comparison.
How offline compares to the built-in Apple Dictation
Apple Dictation is available on every Mac but requires a network connection to function — your audio is sent to Apple's servers for transcription. For users who need dictation to work without the internet, that rules it out. Coii VoiceInput offers the same result — words on the screen, unedited — with the added capability of working everywhere, offline or on a flight.
Performance and accuracy without the internet
Dictation speed and accuracy are identical whether the app has a network connection or not. The transcription model runs on your machine, so performance depends only on your Mac's processor and memory, not on server load or latency. If your Mac is fast enough to run the model at all, offline dictation is as fast as online dictation.
This is different from most cloud-based systems, where network latency and server load both affect how long transcription takes. Offline users sometimes see faster results because there is no round-trip, but the model is the same, the output is the same, and you get the accuracy you paid for without server-side processing that might introduce different errors. Speech recognition happens entirely on your machine.
Syncing your setup across machines (without the cloud)
If you activate on multiple machines, each one stores its own model and its own history. Your dictionary is tied to your licence and travels with it when you activate on a new machine, but your history does not sync — it stays on each machine where you created it. This is a trade-off: no cloud sync means no possibility of your data being exposed through the cloud, and no account or credentials to manage. If you want history on a new machine, copy it from the first machine, or paste it into a text file and carry that with you.
Your machine is yours. The data stays on it.
Remote work and locked-down networks
Offline capabilities are essential in secure environments where uploads are forbidden. A SCIF, a classified office, or a healthcare environment where HIPAA compliance requires local-only processing can all use this app without special network rules or approval, because the app never tries to reach the internet after the initial setup.
This also means no firewall exceptions, no proxy configuration, no security review of cloud destinations — the app simply works, because it does not connect.
The model downloads once
When you first open the app after installation, model weights are fetched from a public host and stored on your Mac. This is a one-time download — a few hundred megabytes depending on the language. After it finishes, the model is installed locally and available for offline use indefinitely. Your MacOS updates, your app updates, and your travels to new places do not re-fetch it — it is on the machine.
The download runs in the background and does not block dictation. You can start using the app immediately; if the model is still downloading, the first dictation waits for it to complete, and every dictation after that uses the installed local version.
When you update the app to a new version, the model is replaced if the version changes, which is rare. Most updates are to the surrounding code, not the engine, so the model you have stays installed.
Offline first, everywhere
The architecture of this app is offline-first. Everything is built to run on your machine. Network access is optional — it is used for setup and for a clock check at launch, but none of it is essential to dictation. This is different from most dictation software, which considers the network essential and offline a fallback.
For anyone who spends time without reliable connectivity — frequent travelers, people in remote areas, workers in secure facilities — offline-first means you stop planning around when you will have internet and just use the app wherever you are.