Offline Speech to Text on Windows: What Works

Windows voice typing does not work offline. Microsoft's Win+H voice typing needs an internet connection, so it stops when the connection does ¹. Two routes do work without Wi-Fi. Voice Access works without an internet connection once its initial setup is done ^{2 3}. Open-source apps such as Buzz, Handy and Murmure run Whisper or NVIDIA's Parakeet model on your own computer, so transcription needs no connection. The rest of this guide covers what each route needs, how accurate offline models are, and which one fits your job.

Windows voice typing stops working without an internet connection

Voice typing is the Win+H tool that types into any text box. To use it you need an internet connection, a working microphone and the cursor in a text box ¹. Microsoft's help page for voice typing lists the internet as a requirement and does not mention offline use. If you need typing without Wi-Fi, the built-in voice typing is not the tool. For how to start it and what it does, see our guide to Windows voice typing.

For the direct answer to the question, see whether Windows voice typing works offline.

Voice Access recognizes speech on the device and works offline

Voice Access is the built-in Windows option for working offline. Microsoft describes it as a feature that lets you control your PC and write text by voice, without an internet connection ². It is available in Windows 11 version 22H2 and later ⁴.

The catch is the first run. Microsoft says the initial setup needs an internet connection ³. Setup downloads the language files for on-device recognition, and that download needs the connection ⁵. Do the setup while online, then test it offline. The steps below use the menu names from Microsoft's setup page.

1. Connect the PC to the internet.
2. Select Start, then Settings, then Accessibility, then Speech.
3. Turn on the Voice access switch.
4. Select Agree and continue and wait for the language files to download.
5. Turn off Wi-Fi and try a few commands from Microsoft's command list.

Once it is running, you can open apps, switch windows and edit text by voice. For the full command set, see the Voice Access guide and the command list.

We found no published word error rate for Voice Access, from Microsoft or from an independent tester. Judge its accuracy on your own speech. Voice Access also covers a limited set of languages, so check the list before you rely on it for another language.

Whisper and other open models run locally on an ordinary Windows PC

Whisper is OpenAI's open speech recognition model, and its code and model weights are released under the MIT License ⁶. Both Whisper and NVIDIA's Parakeet can run on a computer you own, which is why the apps in the next section work offline.

The size of the model decides how much memory it needs, and memory is what most often limits an ordinary PC. Check the installed memory in Windows before you choose a model. The smallest Whisper model, tiny, has 39 million parameters and needs about 1 GB of memory ⁷. The small model needs about 2 GB ⁸. The largest, large, has 1,550 million parameters and needs about 10 GB ⁹.

Installing Whisper yourself takes a few terminal commands, so it suits people comfortable with the command line. The Whisper package requires the ffmpeg tool as well as Python ¹⁰. The package itself installs with pip ¹¹. The step-by-step version is in how to install Whisper on Windows.

The whisper.cpp project is a C and C++ implementation of Whisper. It lists CPU-only inference as a supported option ¹², which means it runs without a graphics card. For a wider walkthrough, see how to run Whisper locally.

Offline dictation apps trade a one-time model download for privacy and no network delay

Open-source apps are the practical offline choice for typing into any app. Their main cost is a one-time download of the speech model, which takes disk space and memory. Their main benefit is that your audio is transcribed on your PC, with no trip to a server and no waiting on a network.

Handy calls itself a free, open-source speech-to-text application that works completely offline ¹³. It is listed as working on Windows, macOS and Linux ¹⁴. It also offers a Parakeet V3 model that its README describes as CPU-optimized ¹⁵.

Murmure is an open-source Windows dictation app that runs Parakeet TDT 0.6B v3 on the PC. Its README says that Murmure "turns your voice into text with no internet connection and zero data collection" ¹⁶.

Buzz transcribes and translates audio files offline on your computer ¹⁷, and its README lists Whisper.cpp as one of its backends ¹⁸. It ships a Windows installer from its project page and a pip package. Nvidia GPU support on Windows needs the pip install with a CUDA build of torch ¹⁹.

The nerd-dictation project is often searched for Windows, but its README describes it as offline speech to text for desktop Linux ²⁰. Treat it as a Linux tool. A directory listing also describes a Windows product called Murmur, with accuracy claims we could not verify. It is not the same project as Murmure, and we leave it out of the table.

Offline does not guarantee privacy on its own. An app can be offline for transcription and still check for updates, send crash reports or store your text in a history. Read the privacy policy and release notes of any app before you trust its claims, and watch the network activity of a new app during your first test. For more on that question, see whether dictation is private.

Choose the app by what you need to do. A live dictation app is the right tool when you type into email or a document. A transcription app is the right tool when you already have a recording. The table is a starting point, not a ranking, because each project changes its features between releases.

Your hardware sets offline speed: CPU, GPU and NPU each help in different ways

The processor runs every offline model. A graphics card (GPU) speeds up some models when the app supports it. A neural processing unit (NPU) helps only with software written for it. Choose hardware by the software you plan to run.

The CPU is the baseline. Handy describes its Parakeet V3 model as CPU-optimized ¹⁵, and whisper.cpp runs without a GPU ¹². NVIDIA says the Parakeet TDT 0.6B v3 model is designed or optimized to run on NVIDIA GPU-accelerated systems ²¹, so an NVIDIA GPU is the hardware its maker built it for. For Whisper on an Nvidia GPU in Buzz, the README asks for a CUDA build of torch ¹⁹.

An NPU is a separate processor for AI tasks. Microsoft says many of its new Windows AI features require an NPU that can run at 40 or more trillion operations per second (TOPS) ²². Those are Microsoft's own features. They do not mean every speech app uses the NPU. The whisper.cpp README describes one path: with a dedicated NPU, it can run the Whisper encoder through VitisAI, which the README says gives a significant speedup over CPU-only inference ²³. That path needs specific hardware and a multi-step setup, so it suits advanced users.

Model choice changes speed a lot. The Open ASR Leaderboard compares 86 open-source and proprietary systems across 12 datasets ²⁴ and reports speed as an inverse real-time factor (RTFx). A higher number means faster processing. Its English short-form results for three models are below.

RTFx depends on the hardware each test ran on, so compare the ratios more than the raw numbers. The sources for these three values are the same leaderboard results cited in the accuracy section below ^{25 26 27}. Your PC will give different absolute speeds.

We did not find a published speed test of these apps on a typical Windows laptop. Test your own PC with your own audio.

Offline models are accurate on clear speech, and noise and telephone audio are the known weak spots

On clear speech, offline models are accurate. On the LibriSpeech Clean test set, Whisper Large V2 had a word error rate of 2.7 percent ²⁸. On the Open ASR Leaderboard, NVIDIA's Canary Qwen 2.5B had an average word error rate of 5.63 percent ²⁶. Parakeet TDT 0.6B v2 had 6.05 percent ²⁵, and Whisper Large v3 had 7.44 percent ²⁷. Those averages cover many test sets, not your voice, so treat them as a guide.

Accuracy falls on harder audio. The same Whisper Large V2 model had a word error rate of 13.8 percent on the Switchboard test set ²⁹, and 25.5 percent on CHiME6 ³⁰. Read those numbers as a pattern, not a promise: the same model can be near 3 percent ²⁸ on one test set and above 25 percent ³⁰ on another, so the audio you record matters more than the model's headline score.

A word error rate counts the words a system gets wrong, whether it swaps, drops or invents them, divided by the number of words in a correct transcript. A rate of 5 percent means about one word in twenty is wrong. For how the number is calculated and read, see word error rate.

Whisper can also invent text. A 2024 study of the 2023 version of Whisper found that roughly 1 percent of its transcriptions of recordings from people with and without aphasia contained entire hallucinated phrases or sentences that did not exist in the audio ³¹. Read a transcript before you send or publish it.

We did not find a published test that runs Windows voice typing and an offline model on the same recordings, and we found no published word error rate for voice typing. So the accuracy gap between Windows' cloud dictation and an offline model on your voice is unmeasured. Your own test is the only reliable answer. For more on how accuracy is measured in general, see how accurate speech recognition is.

Pick by job: live dictation, file transcription or full voice control

Choose by the job, not by the brand. Each route has a clear best fit.

Live dictation into a document is where offline apps help most, because they need no network trip. For full control, use Voice Access. It is the built-in route that works offline for both commands and text. For recordings you already have, use Buzz or the Whisper command line, and pick a model size that fits your memory from the chart above.

Yap, which we make, is an on-device dictation app for Mac and Windows, so the same trade-offs apply to it. You can download it and test it offline yourself.

If the simplest route does not work offline on your own speech, try a larger model or a different app, and check the questions below for setup problems.

Test your offline setup in about ten minutes

Test before you rely on any route. A quick test on your own voice tells you more than any benchmark on this page.

1. Finish the setup or model download while the PC is online.
2. Turn off Wi-Fi and Ethernet, and confirm the PC has no connection.
3. Dictate two minutes of your normal writing, such as a short email, into the app you use most.
4. Read the result against what you said and count the wrong words, the missing words and the invented words.
5. Repeat with a noisy background, such as a fan or a second voice, so you see the weak spot before it matters.

If the errors are few, the route works for you. If they are many, switch to a larger model where the app offers one, and move the microphone closer to your mouth before you try again.

Keep going

• Does Windows voice typing work offline?
• How do you install Whisper on Windows?
• Is there a Whisper app for Windows?
• Which open-source dictation apps run on Windows?
• Windows voice typing: how to dictate with Win+H
• Best dictation software for Windows, free and paid
• Windows dictation privacy and online speech recognition
• Open-source dictation software on GitHub
• How to transcribe audio to text
• Dictation on Windows: voice typing, Voice Access and apps