In short
- Offline speech recognition relies on an AI model downloaded once to your drive, then run by your PC. The audio never leaves the machine.
- The graphics card speeds things up because transcription comes down to huge batches of multiplications done in parallel. Without one, the processor takes over, more slowly.
- The trade-offs: an initial download, memory in use, battery drain on a laptop, and a limited set of languages.
- Ondula runs NVIDIA's Canary model locally, in 25 European languages, on a graphics card of any brand or on the processor. This offline dictation is free, unlimited and needs no account.
Offline speech recognition relies on an AI model copied once to your drive, then run by your graphics card or processor: sound comes in through the mic, text comes out at the cursor, and nothing travels over the internet. Once the model is in place, you could unplug the network cable.
You're dictating a report on the train, with no signal between stations, and the text still appears. So where did the "server" doing the work go? There isn't one. Here's what happens inside the machine, without the jargon: what the model is, why the graphics card matters, what running locally costs your PC, and when the cloud is still the better choice.
What is a speech recognition model?
A model is a file full of numbers, called parameters. They were tuned during a long training process: the system was fed an enormous amount of speech along with its transcript, and its numbers were adjusted after every mistake until it guessed right. Nobody reads those numbers directly. Together, they encode how sounds map to words, accents and run-together speech included.
The dictation software itself is just a wrapper around the model. It captures the mic, hands the audio to the model, then pastes the text in the right place. So two apps that look very different can use the same model, and two apps that look identical can hide one model on your drive and another on the far side of the internet.
The whole difference between local and cloud comes down to one question: where does the model live?
How does offline dictation work inside the PC?
Locally, your voice never leaves the machine. In the cloud, it makes a round trip over the internet to the provider's servers. Here's the local version, step by step:
- Once and for all, the app downloads the model. Ondula tells you its size and asks before it does. That's the only time the internet is needed.
- At startup, the model is prepared and loaded into memory: the graphics card's memory if it can be used, otherwise the PC's RAM.
- When you speak, the sound is converted into a kind of picture of your voice's frequencies, a bit like an equalizer display frozen in time. That picture is what the model reads.
- The model works in two stages. The first part "listens": it summarizes what was said. The second part "writes": it produces words one at a time, taking the previous ones into account. That's how it can choose between "write," "right" and "rite" depending on the sentence.
- The text appears as you speak, then is pasted at the cursor, in whatever app is open: an email, a Word document, a form field.
None of these steps needs a server. That's why truly local dictation works on a plane too.
Why does a graphics card speed up speech recognition?
Every stage of the model boils down to multiplying large tables of numbers together, in enormous quantities for each second of speech. A processor has a handful of very versatile cores: it does the job, but in small batches. A graphics card was designed to calculate the color of millions of pixels at once. It lines up thousands of small compute units working in parallel, which is exactly what a speech recognition model needs.
There's one practical hurdle: many local tools can only take advantage of NVIDIA cards. Ondula uses Vulkan, a standard interface understood by NVIDIA, AMD and Intel drivers. It automatically picks the dedicated card if the PC has one, otherwise the graphics chip built into the processor, otherwise the processor alone, and keeps transcribing live in every case, just with more effort from the machine.
Two details matter on a laptop. On battery, Ondula releases the dedicated graphics card fifteen minutes after your last dictation so it can power down; the model then waits in RAM. And Ondula's website says it plainly: on a modest PC, Cloud gives the best quality. Local works everywhere, but it won't turn an old laptop into a workstation.
Whisper, Canary: which model powers local dictation?
If you've searched for "Whisper transcription Windows," you know Whisper, the open speech recognition model released by OpenAI. It kicked off the wave of transcription on PCs. A good share of the small local dictation tools for Windows are wrappers around it, more or less polished, and many ask you to pick a model size yourself: small and fast but rough, or large and accurate but heavy. Our roundup of open-source dictation tools for Windows reviews several of them.
Ondula made a different choice: a single model, NVIDIA's Canary, on every machine. Nothing to pick, nothing to tune. Canary recognizes 25 European languages offline and can also translate while you dictate, from your language into English or from English into the other 24: you speak German, the text comes out in English, still with no connection. How to use it is explained in our article on speech-to-text translation.
Whatever the tool, two questions matter more than the model's name: does it recognize your language offline, and does it keep up with your voice on your machine?
What are the benefits of local speech recognition?
Privacy you can verify. In the cloud, privacy rests on commitments: how long data is kept, where it's processed, which subcontractors are involved. Those can be serious, and we detail ours in our article on voice typing and GDPR. Locally, there's nothing to take on trust: the audio doesn't leave the PC. Notes from a medical appointment, a memo on a client file, a draft resignation letter all stay with you.
Dictation that doesn't care about the network. A tunnel, a plane, a basement meeting room, overloaded hotel Wi-Fi: the local model can't tell the difference. If your question is mostly practical, our guide to Windows voice typing offline compares what actually works offline in Windows and gives you the airplane mode test that settles it.
No per-minute cost. An online service pays for its servers with every second of audio, which is why cloud plans count hours. Locally, your PC does the work, and a minute of dictation costs nobody anything. With Ondula, offline dictation is free, unlimited and needs no account: notes all day, an audio file transcribed in the evening, no meter running. A single dictation can last up to 2 hours offline, versus 30 minutes online.
What are the limits of a local model?
Local isn't free for your machine. Here's what it asks of it.
- A big download up front, and disk space. Ondula shows the exact size before downloading.
- Memory. The model stays loaded so dictation is ready to go the moment you press the shortcut.
- Power. On an unplugged laptop, every sentence puts the graphics card or processor to work.
- Lag on smaller machines. Without a graphics card, on a modest processor, the text may arrive later than it would online. Ondula warns you if it can't find a compatible card and offers to switch to online.
- A fixed set of languages. Offline, that means 25 European languages: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian and Ukrainian.
Local or cloud: which should you choose for dictation?
The two approaches suit different situations. Ondula offers both in the same app, and you switch between them in Settings > Transcription, choosing "Offline" or "Online."
| Local (offline) | Cloud (online) | |
|---|---|---|
| Where the model runs | On your PC (graphics card or processor) | On servers |
| Does the audio leave the PC? | Never | Yes, for the time it takes to transcribe, and it isn't kept afterwards |
| Without internet | Yes, once the model is downloaded | No |
| What it asks of the PC | Compute, memory, power | Almost nothing |
| On a modest PC | Works, more slowly | The best quality |
| Maximum length of a dictation | 2 hours | 30 minutes |
| Cost with Ondula | Free, no limit | €5.99 a month (20 h), Cloud Pro €9.99 a month (40 h) |
In practice:
- Healthcare, law, HR: Local, so your voice never leaves the workstation.
- Frequent travel: Local, which doesn't depend on any connection.
- An old laptop that already struggles: Cloud, which asks almost nothing of it.
- Lots of dictation, no budget: Local.
How to see what local dictation asks of your PC
Windows shows you everything, no extra tool needed.
- Open Task Manager (Ctrl + Shift + Esc), Processes tab.
- Find Ondula and look at the CPU, Memory and GPU columns.
- Dictate a few sentences locally: the numbers rise while you speak, then drop back.
- For the big picture, the Performance tab shows graphs for the processor and each graphics card.
If usage stays maxed out for a long time and the text lags, your machine is struggling, and Cloud will suit it better. If it spikes briefly and then falls back, local is running comfortably.
Ondula in Local mode
Ondula is a voice dictation app for Windows 10 and 11, distributed through the Microsoft Store. One shortcut, you speak, and the text is typed at the cursor, in all your apps.
- Free, no account: unlimited offline dictation, translated dictation, personal dictionary and voice shortcuts, audio and video file transcription, captures and the palette.
- Lifetime Local, €19.99 once, no subscription, on 3 PCs: adds live meetings and captions, speaker recognition and voice memory, all on your machine.
- Cloud (€5.99 a month) and Cloud Pro (€9.99 a month), cancel anytime: online dictation, AI modes and PDF reports. Create an account and you get 14 days of Cloud with 10 hours online, no credit card needed. Details on the pricing page.
Frequently asked questions
Can speech recognition AI really run on an ordinary PC?
Yes. Today's models fit on a typical PC. A graphics card, of any brand, speeds up the work considerably; without one, the processor takes over, more slowly. On a very modest PC, Cloud is the more comfortable option.
Do you need internet for offline speech recognition?
Only to download the model, once. After that, transcription happens entirely on the PC, airplane mode included.
Does Ondula use Whisper?
No. In Local mode, Ondula uses NVIDIA's Canary on every machine. It recognizes 25 European languages and can translate while you dictate, into or out of English.
Is local dictation less accurate than the cloud?
It depends on the model, the language and the machine. An online service can run heavier models than a laptop. The most reliable test is to dictate the same paragraph in both modes, which the free Cloud trial lets you do.
Why is my local dictation slow?
Most often, the model is running on the processor because there's no usable graphics card, or the machine is already busy with other programs. Close resource-hungry apps, check in Task Manager that the GPU is working during dictation, and on a very modest PC, switch to Cloud.
To watch dictation AI run on your own machine, download Ondula for free from the Microsoft Store: offline dictation is free and unlimited, in 25 European languages.
