Fundamentals

Local vs Cloud Speech to Text: How to Choose

Local speech to text runs on your own computer, works offline and keeps the audio on the device; cloud speech to text sends the recording to a server, where larger models can transcribe more languages and clean up the result. Choose local for privacy, offline work and zero running cost, and cloud for the most polished text or for languages and hardware that local models do not cover.

You do not always have to pick one for good. AirTypes lets you choose each time: Plain is free and offline, with Whisper running on your own CPU, and Polished cleans up your speech in the cloud at 120 minutes for every dollar. Here is the same sentence through both:

You say

"ok so um the meeting is moving to tomorrow at ten, client is not available, and uh send him the invoice again"

→
Plain types (local)

"The meeting is moving to tomorrow at ten, client is not available, and uh send him the invoice again."

You say

"ok so um the meeting is moving to tomorrow at ten, client is not available, and uh send him the invoice again"

→
Polished types (cloud)

"The meeting is moving to tomorrow at 10:00. The client is not available. Send him the invoice again."

Plain gives you your own words with light cleanup: the leading "um" goes, the first letter is capitalized and a period is added. Polished removes the filler in the middle too and splits the thought into clean sentences. Both are useful; they suit different moments.

How local and cloud speech to text work

Every speech to text system does the same basic job: it turns a recording into a spectrogram, runs it through a trained model and writes out text. The difference is where that model runs.

Local (also called on-device or offline speech recognition) downloads the model to your computer once. From then on, your microphone, the model and the text all stay on the machine. OpenAI's open Whisper models made this practical on ordinary laptops; tools such as whisper.cpp, Superwhisper's free local models and AirTypes Plain are built on that family. Our explainer on offline speech recognition walks through the pipeline step by step.

Cloud (online) speech recognition records on your device and uploads the audio to the provider's servers. A model there transcribes it, often a far larger one than a laptop could run, and the text comes back. Many services then pass the text through a language model to tidy it. Wispr Flow, for example, always transcribes in the cloud, and Otter.ai is a cloud meeting transcription service.

Privacy

This is the clearest difference. With local speech to text, the audio does not need to go anywhere. You can check it yourself: turn off Wi-Fi and dictate. If it still works, the recognition is happening on your machine. AirTypes Plain uploads 0 bytes of audio and works in airplane mode.

With cloud speech to text, your recording leaves the device by definition. That is not automatically a problem, but it means the provider's policy now matters: how long recordings are kept, whether they are used to train models, who can access them, and where the servers are. Some services offer extra controls, such as Wispr Flow's Privacy Mode. Read the policy of any cloud tool before you dictate confidential material.

Be precise about what "local" covers. Even with on-device transcription, once you paste the text into an email or a chat app, that app handles it as normal. Local speech to text protects the voice step, not everything that happens after. Our guide is voice dictation private? lists the questions worth asking any tool.

Accuracy and cleanup

Local models are limited by your hardware. A laptop can comfortably run a small or medium Whisper model, and a well-equipped one can run a large one. AirTypes Plain keeps to the two smallest English models, which it calls Swift and Crisp, so it runs on a 4 GB machine. For clear English in a quiet room, even the smaller models produce good text. They struggle more with heavy accents, background noise and unusual names, and the smallest ones make more mistakes overall. See the Whisper model sizes guide for the trade-off at each size.

Cloud services are not bound by your laptop, so they can run the largest speech models and add an AI cleanup step on top. That second step is the bigger difference in practice. Spoken language is messy: "um", repeated words, false starts, sentences that never end. A cloud language model can remove the noise and fix grammar and punctuation, so the text reads as if you had typed it.

You say

"the results suggest that um the effect is, the effect is stronger in the treatment group, which is consistent with uh Smith twenty nineteen"

→
Polished types (cloud)

"The results suggest that the effect is stronger in the treatment group, which is consistent with Smith 2019."

The risk with AI cleanup is that it changes your meaning. A good system guards against that. In AirTypes Polished, if the rewrite drops a number you said or drifts too far from your length, you get your own words instead. It is a mechanical check, not a reader, so skim anything important. It can also drop a greeting such as "hi" at the start of a message, which is worth knowing if you dictate replies.

Languages

The Whisper family includes multilingual models, so local transcription in other languages is possible in principle. In practice, the models small enough for an average laptop are weaker outside English, and many local tools focus on English. AirTypes Plain is English only.

Cloud models have more room for multilingual work, and they can translate as they go. AirTypes Polished works in the language you speak. That includes English, Hindi, Bangla, Japanese, German and most other spoken languages, so try yours. With the optional Always translate to English setting, you speak your language and English is typed:

You say (Bangla)

"কাল সকালে মিটিং টা হবে না, client এর flight delay হয়েছে, আমি নতুন time টা email করে দিচ্ছি"

→
Polished types (Always translate to English)

"The meeting will not be tomorrow morning. The client's flight has been delayed. I am emailing the new time."

Cost and hardware

Local speech to text costs nothing to run once the model is on your machine. You pay with hardware instead: RAM to hold the model and CPU time to run it. Small Whisper models run on a 4 GB machine with no GPU; the larger, more accurate ones need more memory and take longer to return text. If your laptop is older, our guide to voice to text on a slow laptop covers which model to pick.

Cloud speech to text moves the hardware cost to the provider's servers, and somebody pays for those. Most cloud dictation apps charge a monthly subscription, often with a free plan capped by words or minutes. Some charge per use. The upside is that a cheap or old laptop can produce the same quality as a new one, because it only records and types.

AirTypes Plain is free. Polished uses prepaid credit: $5 buys 600 minutes of speech, charged by the second, with one price for transcription and cleanup together. Credit never expires and nothing renews. We compare pricing models across apps in dictation apps without a subscription.

Connectivity and speed

Local speech to text works anywhere: on a plane, on a train with patchy signal, in a building that blocks outside traffic. Its speed depends only on your computer. On a capable machine with a model that fits, text comes back quickly; on a weak one with a model that is too large, you wait.

Cloud speech to text needs a working connection. Each recording has to upload and the result has to come back, so a slow or unreliable network adds delay, and no network means no transcription. On a good connection, the server's speed can more than make up for the round trip, especially compared with a large model on an old laptop.

Local when you want it, cloud when you need it

Plain transcribes offline on your CPU, free. Polished sends the recording for AI cleanup when you want finished text. macOS and Linux.

Download for macOS and Linux

Decision table

Pick the column that wins on the factors that matter most to you.

FactorLocal (on-device)Cloud (online)
Audio leaves your computerNoYes, to the provider
Works offlineYesNo
Running costFree once downloadedSubscription or pay per use
Hardware neededEnough RAM and CPU for the modelAny machine that can record
Accuracy on clear EnglishGood, better with larger modelsLarge models available
AI cleanup of filler and grammarLight or none in many toolsYes, in many services
Other languages and translationLimited on small modelsBroader
Depends on network speedNoYes
  • Choose local if you handle confidential material, work offline, want no running cost, or dictate mostly in clear English on a reasonably capable machine.
  • Choose cloud if you want finished, punctuated text without editing, dictate in a language other than English, or your laptop struggles with local models.
  • Choose both if your work varies through the day. Most people's does.

How AirTypes lets you pick

AirTypes is a desktop dictation app for macOS and Linux, with a browser option described at the end of this section. In the app, you hold Ctrl + Shift + Space, speak, let go, and the text is typed at your cursor in any desktop app. Two modes sit on the Home screen, and you can switch whenever you like, so each recording goes the way you choose.

Plain: local

  • OpenAI's Whisper runs offline on your own CPU. 0 bytes of audio are uploaded, and it works in airplane mode.
  • English only, with two model tiers: Swift (about 78 MB) and Crisp (about 148 MB), the one we recommend.
  • Light cleanup on the machine: a leading "um" or "uh" removed, the first letter capitalized, a period added.
  • Free, with no time limit.

Polished: cloud

  • Polished sends the recording to AirTypes to transcribe and clean it up.
  • A large speech model transcribes it, then an AI removes filler words and false starts and fixes grammar and punctuation, keeping your meaning, numbers and names. If the rewrite drops a number you said or drifts too far from your length, you get your own words instead.
  • Works in the language you speak, with optional Always translate to English.
  • Needs an internet connection. Prepaid credit: $5 buys 600 minutes of speech. When credit runs out, Polished pauses and Plain keeps working, free.

In both modes, AirTypes receives your account email and usage counts. In Plain, nothing else. In Polished, it also receives the recording. A practical pattern: keep Plain on for anything sensitive and for quick notes, and switch to Polished for the emails, reports and messages where clean text saves you an editing pass. The Polished section of the homepage has more examples, and how to use AirTypes covers setup.

On Windows? The Windows desktop app is in development, and you can join the waitlist. Today you can Dictate in your browser (new in November 2026): sign in on airtypes.com, click the mic and speak, and the Polished text is ready to copy and paste. That is the cloud side only: the recording is sent to AirTypes, and Plain, the local mode, needs the macOS or Linux app.

FAQ

Is local speech to text more private than cloud?

Yes, for the voice step. Local speech to text transcribes on your device, so the audio does not need to leave it. Cloud speech to text uploads the recording, and what happens to it depends on the provider's policy.

Is cloud speech recognition more accurate than offline?

Often, because servers can run larger models and add AI cleanup. For clear English, a good local Whisper model is accurate enough for everyday writing. The gap is widest with accents, noise, unusual names and languages other than English.

Does on-device dictation work without internet?

Yes, once the model is downloaded. AirTypes Plain works in airplane mode. Cloud modes, including AirTypes Polished, need a connection.

Can I use local and cloud dictation in the same app?

In AirTypes, yes. Plain runs locally and free; Polished sends the recording to AirTypes for transcription and cleanup. You switch between them on the Home screen.

Which is cheaper, local or cloud speech to text?

Local costs nothing to run once the model is on your machine. Cloud costs money because servers do the work, usually as a subscription or pay per use. AirTypes Polished is pay as you go, 120 minutes of speech for every dollar, with no subscription.

Get AirTypes, free

Plain transcribes offline on your CPU, free with no time limit. Switch to Polished when you want clean, finished text in the language you speak.

Download for macOS and Linux