Offline dictation, and why it matters

What "cloud dictation" actually does

When you dictate into a typical tool, this is the sequence: your microphone is recorded, the audio is uploaded to a server, a model there converts it to text, and the text is sent back. It works well and it is why those tools can be free.

The dashed line is the edge of your computer. Both lanes are drawn on the same clock deliberately — this shows where your audio goes, not how fast anything is.

The part worth understanding is what the upload implies:

• Your raw voice — not just the words — leaves your machine.
• It is processed on hardware you do not control, under terms that can change.
• It may be retained, logged, reviewed for quality, or used to improve models, depending on the provider and the plan you are on.
• It stops working entirely when you are offline.
• Whoever runs the network in between can see that the traffic happened.

For a shopping list, none of this matters. For other things, it decides the question by itself.

When it stops being a preference

There is a category of work where sending audio to a third party is not a trade-off to weigh but a line you cannot cross:

InkBeepAI is software, not a compliance programme. Whether any particular tool fits your obligations is an assessment you or your organisation has to make. What we can do is tell you precisely what the software does — which is the input that assessment needs.

What InkBeepAI does instead

InkBeepAI runs speech recognition on your own PC. Not "mostly on-device", not "on-device where possible" — there is no server in the transcription path at all, because there is no code in the app that would send audio anywhere.

• Both speech models ship inside the download. Together they are about 690 MB of the roughly 910 MB install (141 MB for English, 547 MB for every other language), and both are present before you first run the app. Nothing is fetched at runtime, on first use, or ever.
• Audio lives in memory and is discarded. It is captured while you are dictating, converted to text, and dropped. It is never written to disk, so there is no cache, no temp file and no recording to find later.
• Transcripts are kept out of Windows Clipboard History. InkBeepAI inserts text via the clipboard, so it explicitly tags the data to be excluded from clipboard history and from cloud clipboard sync. Without that, everything you ever dictated would land in your Microsoft account through a side door — this is the most easily missed leak in a tool of this kind, and it is closed.
• Logs record timings, never content. The app logs how long a transcription took and how many characters resulted. It does not log the text.
• The network is used for your licence, never for your speech. InkBeepAI talks to one server and nothing else: once when you activate your key, and then a licence check every two weeks. Each of those carries your licence key and up to three anonymous machine identifiers (one-way hardware hashes). No audio. No transcripts. No account, because there is no account.

How to verify it yourself

You should not have to take a vendor's word for this, and with InkBeepAI you do not have to. Two tests, neither of which requires any technical skill:

1. Disconnect from the internet. Turn off Wi-Fi, unplug the cable, enable airplane mode. Then dictate. Everything works exactly as before — same speed, same accuracy. A cloud tool cannot do this.
2. Watch the network. Use any traffic monitor you like and dictate for an hour. You will see no traffic from InkBeepAI at all, apart from the fortnightly licence check.

If you want to go further, every release is published with its SHA-256 checksum, so you can confirm the file you downloaded is byte-for-byte the one we published.

The trade-offs, stated plainly

On-device recognition is not free of cost. The honest list:

• The download is large — most of it is the two speech models, which is the price of not needing a server. Both ship inside it, so no language ever needs a download.
• It needs a reasonably modern processor. InkBeepAI requires AVX2 support: Intel Core 4th generation (2013) or newer, or AMD Ryzen / FX "Excavator" (2015) or newer. It checks at startup and tells you rather than misbehaving.
• Languages other than English need a graphics processor. Without one they are far too slow to dictate with. InkBeepAI tells you, and offers to switch back to English, which runs on any supported PC.
• 16 languages are verified, not 99. A cloud service can list a hundred languages because the models live on their hardware and claiming them costs it nothing. Ours are on your machine, so we measured what we could on recordings of native speakers: 16 passed and carry a published error rate, 9 more missed the bar and are named with their figures, and 83 are not verified. The table.
• It uses your own hardware rather than someone else's: the graphics processor where it can, otherwise the processor. That work happens while it transcribes, and once more to warm up whenever it loads a speech model.

We would rather list these here than have you find them after paying.

Common questions