Offline speech to text, explained

What an offline speech-to-text desktop app actually is

A desktop app, not a browser tab. That distinction does most of the work. A web-based dictation tool runs in your browser and almost always sends the audio to a server. A real offline app is native software that loads a speech model into memory and runs it on your CPU or GPU, right there on the machine.

The behaviour is simple on purpose. You press a global hotkey from inside any application (your email client, a code editor, a Slack box, a clinical note field) and the transcript appears where your cursor is, like a fast typist sitting in for you. No copy-paste from a separate window. No "upload your recording." It behaves like the keyboard, because that is the only interface every app already understands.

The default hotkey in Whisper is Ctrl+Space on Windows and Linux. On a Mac you hold Command and Option together and release either key to stop, a push-to-talk chord, not a toggle. You can rebind it; the first thing a lot of people do is move it off whatever conflicts with their drawing app or their music software. I hardcoded that hotkey in an early version and got a 2 a.m. email about it, which is a story for the FAQ.

The honest version of "how" is unglamorous: there is no magic, just a model file on your disk. When you install a local app, you download a speech model, a few hundred megabytes to a few gigabytes of trained weights. From then on, transcription is a math operation the model performs on your audio. The audio never becomes a network request. Whisper by Remskill runs all of its local transcription through a pure-Rust library called transcribe-rs, with no Python sidecar and no separate runtime to keep alive. The thing you feel is the app starting fast and the audio path staying short.

Press a key, speak, get text — the offline loop

Here is the full loop. You hold the hotkey. A small overlay tells you it is listening. You talk, a sentence, a paragraph, a 600-word lecture summary. You let go. The model transcribes what you said and the text lands at your cursor, not in a file you then have to go find. That is the practical difference between a dictation app and a file-transcription tool: a dictation app inserts text live, the moment you stop talking.

None of that needs a network. The microphone feeds audio into the local model, the model produces text, and the app types it for you. The recording continues for a short tail after you release the key, so the last word does not get clipped, a small thing that took me longer to get right than I would like to admit.

For everyday email and notes, the round trip from key-release to pasted text is a second or two with a Small model on a recent laptop, fast enough to stay in the sentence you were thinking. There is an optional second step too: Whisper can run an AI cleanup pass (fixing punctuation, formatting a list, or answering a question you spoke), and in local mode that pass runs on-device through a local engine called Ollama. So even the "make this sound less like spoken rambling" step can stay entirely offline. You trigger it by starting your sentence with "Hey whisper". Ignore it and you get a clean verbatim transcript. Either way, nothing you say leaves the machine unless you flip the optional Cloud switch.

Parakeet or local Whisper: which engine runs offline best

Two local engines ship in the app, and they are good at different things. This is the one real decision you have to make.

Parakeet is NVIDIA's TDT model, about 600 MB, and it is the fast one, 5 to 10 times faster than Whisper on a CPU. It handles English plus 24 European languages, 25 in total. It does not translate to English, and it has no custom-vocabulary controls. If you mostly speak English and you want speed, this is the one.

Whisper is the multilingual workhorse. Its multilingual models support 99 languages with auto-detect and can translate any of them to English. It is slower than Parakeet on the same hardware, but it gives you Chinese, Japanese, Korean, custom vocabulary, and finer control. One thing to keep straight: the English-only Whisper models (the ones labelled with .en) lock to English alone and skip the 99-language support entirely. So "Whisper does 99 languages" is only true for the multilingual builds.

Rule of thumb: English and in a hurry, pick Parakeet. Anything multilingual, anything you need translated, pick a multilingual Whisper model.

Picking the right local model for your hardware

A local model loads into RAM, so your hardware sets the ceiling. The good news for most people: the smaller models are plenty.

On the English side, the models run from Base at about 140 MB, to Small at about 480 MB, to Medium at about 1.5 GB. On the multilingual side, you go from Small at 480 MB up to Large v3 at about 3 GB. The Small multilingual model is what new users start on, and it is a sensible default: it fits anywhere and it is accurate enough for everyday dictation.

The boring truth about hardware is that the microphone matters more than the model. A $20 USB mic does more for accuracy than jumping two model sizes up. I spent a weekend convinced a model upgrade would fix my muddy transcripts; it was the laptop mic picking up the dishwasher the whole time. If you are on an older laptop with 8 GB of RAM, run Parakeet or a Small model and spend the upgrade money on the mic. If you have a recent Apple Silicon Mac or a 16 GB Windows machine with a discrete GPU, every model runs comfortably, including Large v3. You do not need a GPU to start: Parakeet on a CPU is the fast path because it was built for that.

Why offline beats the cloud for sensitive audio

Cloud-only dictation is a privacy disaster waiting to be transcribed. I will say that as a flat opinion and then back it, because hand-waving about privacy is cheap.

A team I worked with once had a contractor build an internal "AI dictation" prototype on top of a cloud model. It called the API for every utterance, on every laptop, with a "smart retry" that re-sent recordings when it got nervous. At the end of the quarter the manager opened the cloud-cost dashboard and found a five-figure bill, most of it from one team transcribing standup recordings four times over. The contractor's fix was "we should optimize the prompt." The CFO's fix was "or we should not pay for cloud transcription of meetings that already have notes." Two problems in one bill: the cost, and the fact that every word of those meetings now lived in a vendor's logs.

Offline solves both. Your boss's salary spreadsheet, the email to your kid's school, the legal brief you are drafting on a train with no signal: none of that should leave your laptop because you wanted to type with your voice. The machine already has a microphone and a processor. For one paragraph, it does not need a server in the loop. Offline also works with the Wi-Fi off, which the cloud, by definition, cannot do.

When the cloud or a built-in tool is the better call

Here is the part the single-tool pages skip. A local app is not always the right answer.

First, the question people keep asking: does Windows do speech to text offline? Mostly no, for the default tool. Windows voice typing, opened with Win+H, requires an internet connection per Microsoft's own documentation, so the dictation is cloud-based. A separate Windows 11 feature called Voice Access does work offline once its language files are installed, but it is built for controlling the PC by voice, not fast longform dictation. On a Mac, Apple's built-in Dictation can process some general text on-device, but Apple's wording hedges: it is not guaranteed for every language, and some dictation may still go to Apple's servers. So if all you ever dictate is the occasional 30-word reply, the built-in tools are free and fine for short bursts. A dedicated app earns its place once you are dictating real volume (emails, notes, documents), not one-line texts.

Second, the cloud wins for some jobs. If you need the highest possible accuracy on hard audio (heavy accents, overlapping speakers, terrible mic conditions), a cloud model can be more accurate than a local one, and it can pull live web answers a local model has no way to know. Whisper's optional Cloud mode uses OpenAI through your own API key for exactly those cases, listing roughly 57 languages. The trade-off is the honest one: it needs internet, and it sends your audio to OpenAI. If your work is not sensitive and you want the best quality plus web search in one place, the cloud is the better pick, and I would tell you to use it. If your audio is private or your connection is unreliable, stay local.

If you have been eyeing a cloud dictation tool but want your audio to stay on your machine, our free Aqua Voice alternative makes the offline, private case in full.

Offline, online, or built-in: a quick comparison

A short table to settle the choice. The fuller breakdown of paid dictation tools lives in our transcription software guide.

Whisper ships the first three in one app: the local pipeline is free, and Cloud is the optional add-on. If you came here for "private and works on a plane," the first two rows are your answer.

If you take one thing from this

Offline speech to text is not a compromise you settle for when the Wi-Fi is down. It is the version where the audio is yours, the latency is short, and there is no quarterly cloud bill waiting at the end of it. The cloud still wins for the hardest audio and live web answers, and I will point you there when it fits. For everything else (the email, the note, the brief on the train) the model on your own disk is enough, and it has been for a while now.