Fundamentals

Voice to Text for a Slow Laptop: Lightweight Dictation That Works

Voice to text works on a slow laptop if you match the speech model to the machine: a small Whisper model runs on 4 GB of RAM with no GPU. If even that feels heavy, send the recording to a server instead, so your laptop only records and types.

AirTypes gives you both routes in one app. Plain is free and offline: Whisper runs on your own CPU, and both of its tiers are small enough for modest hardware. Polished cleans up your speech with AI at 120 minutes for every dollar, and because the heavy work happens on our servers, it is easy on a weak laptop. You switch between them on the app's Home screen.

Even the small local models give you usable text. Here is a real Plain result, transcribed on the machine with its light local cleanup:

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 (offline)

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

What makes local speech recognition heavy

When speech recognition runs on your computer, your computer does all the maths. Four things decide how hard that is.

Model size

A speech model is a large file of learned numbers. OpenAI's Whisper comes in several sizes, from tiny (about 39 million parameters) up to large-v3 (about 1.55 billion). Every second of audio has to pass through the whole network, so a bigger model does far more work for the same sentence. Larger models make fewer mistakes on unusual words and unclear audio; smaller ones are much lighter. Our Whisper model sizes guide covers the full range.

RAM

The model has to sit in memory while it works, along with the audio and some working space. On a laptop with 4 GB of RAM, the operating system and a web browser already use a good share of that. Load a model that does not fit comfortably and the system starts swapping memory to disk, which slows everything down, not just dictation.

CPU cores and threads

Whisper splits its work across CPU cores. A dual-core laptop from several years ago has far fewer hands for the job than a recent six- or eight-core chip, so the same model takes noticeably longer. Laptops also slow their processors down when they run hot or on battery, which matters on older machines with tired fans.

No GPU required

A graphics card can speed up the larger models, but it is not needed for the small ones. Whisper runs on an ordinary CPU, which is why speech to text on an old computer without a dedicated GPU is perfectly practical, as long as you choose a model that suits it. For the bigger picture of how this works without the cloud, see how offline speech recognition works.

Which Plain tier to pick for your hardware

AirTypes Plain has two tiers, both small Whisper models. Each downloads once and then stays on your machine. Both run on the CPU, handle English and upload no audio. Crisp is the tier we recommend; Swift is the one to use when Crisp feels slow.

TierWhisper modelDownloadRough starting pointLight enough for an old laptop
Swifttiny~78 MB4 GB RAMYes
Crispbase~148 MB4 GB RAMYes

4 GB of RAM or an older dual-core chip

Start with Crisp, and switch to Swift if you wait longer than you would like for the text. Swift is the lightest and fastest; Crisp makes fewer mistakes for a little more work. For notes, chat replies and short emails in clear English, either is a sensible daily choice. Speak in full sentences and keep background noise down; small models gain more from a clean recording than large ones.

More than 4 GB and a reasonably recent CPU

Use Crisp. It is the larger of the two Plain tiers, so there is no bigger local model to move up to. If you need more accuracy than Crisp on jargon, accents or long sentences, Polished is the route, whatever hardware you have.

On an older Intel MacBook, download the Universal build and start on Crisp; the macOS setup guide walks through the install.

Signs the model is too big

You do not need a benchmark to know a model is too heavy for your machine. Watch for these:

  • A long pause after you let go of the hotkey before any text appears, especially on short sentences.
  • Fans spin up hard every time you dictate, or the laptop gets hot quickly.
  • Other apps stutter while transcription runs: the browser freezes, music skips, the cursor lags.
  • Heavy disk activity or a full swap file, a sign the system has run out of RAM.
  • The battery drains noticeably faster on days you dictate a lot.

If you see two or more of these on Crisp, switch to Swift and try again. A smaller model that answers quickly beats a larger one you avoid using because it makes the laptop crawl.

Dictation that fits the computer you have

Plain runs small Whisper models offline on your CPU, free. Polished does the heavy work on our servers when your laptop cannot. macOS and Linux.

Download for macOS and Linux

Get more from an old computer

A few habits help any local speech model run better on modest hardware:

  • Close what you are not using. Browser tabs and chat apps hold a surprising amount of memory. Freeing some lets the model stay in RAM instead of spilling to disk.
  • Plug in for long sessions. Many laptops run their processor slower on battery to save power.
  • Keep the vents clear. A laptop that overheats slows itself down. A hard surface helps more than a blanket or a cushion.
  • Dictate in shorter bursts. A few sentences at a time keeps each transcription small and gives you text sooner.
  • Use a decent microphone. A headset or a USB mic gives the model cleaner audio, which helps small models most.

On Linux, a lightweight desktop environment also leaves more memory for the model. Our Linux dictation guide covers the Linux setup, and how to use AirTypes walks through choosing a tier on first launch.

Skip the local load with Polished

Sometimes the honest answer is that the laptop is too old for the accuracy you want. Polished takes the local load off entirely. Polished sends the recording to AirTypes to transcribe and clean it up: a large speech model transcribes it on our servers, then an AI removes filler words and false starts and fixes grammar and punctuation, keeping your meaning, numbers and names. Your laptop only records and types the result.

You say

"yesterday I am going to client office and they are telling the payment will be late by two week, so we need to discuss about the budget again"

→
Polished types

"Yesterday, I went to the client's office, and they said the payment will be late by two weeks. So, we need to discuss the budget again."

What to know before you switch:

  • It needs an internet connection. Plain is the one that works offline.
  • It is paid, from prepaid credit. $5 buys 600 minutes of speech, charged by the second. Credit never expires and nothing renews. When it runs out, Polished pauses and Plain keeps working, free.
  • It works in the language you speak. That includes English, Hindi, Bangla, Japanese, German and most other spoken languages, so try yours. An optional Always translate to English setting types English whatever you speak.
  • If the rewrite drops a number or drifts too far from what you said, you get your own words instead.

A mix works well on older machines: Plain on Swift or Crisp for quick notes, and Polished for the emails and documents where accuracy and clean punctuation matter. More on the trade-offs in our local vs cloud speech to text guide, and on what pay-as-you-go costs compared with other apps in dictation apps without a subscription. Pricing is on the homepage.

Other lightweight options

Built-in tools such as Apple Dictation, Windows voice typing and Google Docs voice typing cost nothing and put little load on your machine. If they are accurate enough for you, they are a reasonable choice. Open source tools such as whisper.cpp and nerd-dictation can run small models on Linux too, with a manual setup. Cloud-only apps like Wispr Flow also keep the heavy work off your laptop, on a subscription; we compare them in our Wispr Flow alternatives post.

FAQ

Can I run voice to text on a laptop with 4 GB of RAM?

Yes. AirTypes Plain's two tiers, Swift and Crisp, run on a 4 GB machine with no GPU. Close heavy apps while you dictate so the model stays in memory.

Do I need a GPU for speech recognition?

No. A GPU speeds up the larger Whisper models, but the small ones run well on an ordinary CPU. AirTypes Plain runs on the CPU.

What is the most lightweight speech to text option?

For offline use, the smallest Whisper model, which AirTypes calls Swift (about 78 MB). If you want no local load at all, use a server-based mode such as AirTypes Polished, which needs an internet connection.

Why is dictation slow on my old computer?

Usually the model is too large for the RAM or CPU. Switch from Crisp to Swift, close memory-hungry apps and plug in the charger. If it is still slow, switch to Polished so the transcription runs on a server.

Is AirTypes free on an old laptop?

Plain is free on any supported machine, with no time limit. Polished is optional and uses prepaid credit at 120 minutes of speech for every dollar. The desktop app runs on macOS and Linux; the Windows app is in development, and you can join the waitlist. On a Windows laptop today, you can Dictate in your browser (new in November 2026): it is Polished only, so nothing heavy runs on the laptop, and the text is copied for you to paste.

Get AirTypes, free

Plain transcribes offline on your CPU, free, with tiers small enough for older laptops. Polished is there when you want clean text without the local load.

Download for macOS and Linux