guide: local ai dictation on windows

Local AI dictation on Windows: speech and cleanup, both on your PC.

AI dictation does two jobs: a speech model turns your voice into text, then a language model drops the filler words or rewrites the text into an email or a list. An app can run the first on your PC and still send the second to a server. This guide covers four ways to keep both on Windows, the free ones first, and puts Voxmelt, our own app, last. Every outside fact links the page it comes from, with the date we checked it.

the short answer

Four routes, the free ones first

1. Fluid dictation in Windows

Built in on supported Copilot+ PCs: small language models on the PC fix grammar, punctuation and filler words as you speak, in English.

Source: Use Voice Access to navigate Windows and applications, Microsoft Learn. Checked on 27 October 2026.

Read route 1

2. OpenWhispr

Free and open source, with local speech and AI models if you pick them. Its cloud is the default once you finish onboarding.

Sources: OpenWhispr, README, GitHub, OpenWhispr/openwhispr; OpenWhispr, Network Allowlist, GitHub, OpenWhispr/openwhispr. Checked on 27 October 2026.

Read route 2

3. Build your own

whisper.cpp for the speech, Ollama for the AI step, both free and MIT licensed. You write the glue.

Sources: whisper.cpp, README, GitHub, ggml-org/whisper.cpp; Ollama, README, GitHub, ollama/ollama; Ollama, MIT License, GitHub, ollama/ollama. Checked on 27 October 2026.

Read route 3

4. Voxmelt

Both steps in one Windows app, the AI step on an NVIDIA GPU. Free tier, then $49 once.

Read route 4
before you pick

Check both halves, not just the speech

Local transcription is only half of it. The AI step is a second model, and it can run somewhere else. Four checks settle it for any app, ours included.

1. Which speech model, and where does it run?

Look for a model you download, such as Whisper or Parakeet, set as the one in use. If transcription goes through a provider you sign in to or add a key for, your audio goes to that provider.

2. Which language model does the cleanup?

The cleanup is a second model. Look for a local runtime, such as Ollama or llama.cpp, and a model on your own disk. If the AI step uses a cloud model, your text goes to that provider for that step, even when your audio stayed home.

3. What is the default?

“Local or cloud” can still mean cloud until you change it. OpenWhispr’s own network notes, for example, list its cloud as the default once you finish onboarding.

Source: OpenWhispr, Network Allowlist, GitHub, OpenWhispr/openwhispr. Checked on 27 October 2026.

4. Does it work with the Wi-Fi off?

Once the models have downloaded, turn the network off, dictate a paragraph and run the cleanup. If both still work, both run on your PC.

The network test step by step, with what to watch in Windows, is on how to prove a dictation app is really offline.

your hardware

What each half needs from your PC

The speech step: a CPU is enough

whisper.cpp supports CPU-only inference, and Voxmelt’s default engine, NVIDIA’s Parakeet v3, runs on any modern 4-core x64 CPU with no graphics card. Whisper can use a GPU too: whisper.cpp lists NVIDIA, Vulkan and AMD ROCm, and OpenWhispr’s README says its local Whisper uses CUDA, or AMD and Intel GPUs through Vulkan.

Sources: whisper.cpp, README, GitHub, ggml-org/whisper.cpp; OpenWhispr, README, GitHub, OpenWhispr/openwhispr. Checked on 27 October 2026.

The AI step: the heavy half

Windows’ fluid dictation needs a supported Copilot+ PC. Voxmelt’s AI step needs an NVIDIA GPU with 6 GB of video memory or more (Ampere or newer) and suggests 8 to 12 GB or more; with an AMD or Intel card, its dictation still works and the AI step does not. The GPU fit check on the download page shows whether a given model fits on your card.

Source: Understand voice-powered productivity on Copilot+ PCs, Microsoft Learn. Checked on 27 October 2026.

route 1 of 4

Fluid dictation: built into Windows on Copilot+ PCs

If you have a supported Copilot+ PC and dictate in English, try this first: it costs nothing extra, and its language models run on the PC.

Microsoft describes fluid dictation as a feature of supported Copilot+ PCs that suggests corrections as you speak, adjusting grammar, punctuation and spelling in real time to make dictated text more readable.

Source: Understand voice-powered productivity on Copilot+ PCs, Microsoft Learn. Checked on 27 October 2026.

It works by running small language models directly on your Copilot+ PC, which correct grammar, add punctuation and remove filler words as you speak in English.

Source: Use Voice Access to navigate Windows and applications, Microsoft Learn. Checked on 27 October 2026.

Microsoft’s Windows Insider notes of 15 May 2026 say you can try it from voice access or from voice typing (Win+H), and that it was then available in Spanish and French for Windows Insiders.

Source: Windows 11 Insider Experimental Preview Build 26300.8493, Windows Insider Program, Microsoft Learn. Checked on 27 October 2026.

Availability varies by device, language, region, release status and installed Windows updates, and some of it works only in English. Microsoft also asks you to review what fluid dictation writes before you share it.

Source: Understand voice-powered productivity on Copilot+ PCs, Microsoft Learn. Checked on 27 October 2026.

What Microsoft describes is correction as you speak. If you want more than that, the routes below put a general language model on your text. Whether the speech step itself stays on the PC depends on which of the two you dictate with; our guide to private, offline dictation on Windows 11 covers voice access and voice typing, with Microsoft’s pages linked.

route 2 of 4

OpenWhispr: free and open source, local if you choose

A free desktop dictation app that can run both the speech and the AI step on your PC, once you pick local models for both.

OpenWhispr calls itself the open-source and free alternative to WisprFlow and Granola. It runs on macOS, Windows and Linux, with a Windows installer, and its code is under the MIT licence, free for personal and commercial use.

Source: OpenWhispr, README, GitHub, OpenWhispr/openwhispr. Checked on 27 October 2026.

For the speech step it offers local models, Whisper and NVIDIA Parakeet among them, with which, its README says, your audio never leaves your device. It also offers cloud processing for speed.

Source: OpenWhispr, README, GitHub, OpenWhispr/openwhispr. Checked on 27 October 2026.

For the AI step its README says every core feature, AI reasoning included, works with local models or cloud providers, and it credits llama.cpp for local text processing. Its network notes say a local reasoning model downloads only when you opt into one.

Sources: OpenWhispr, README, GitHub, OpenWhispr/openwhispr; OpenWhispr, Network Allowlist, GitHub, OpenWhispr/openwhispr. Checked on 27 October 2026.

The same network notes list OpenWhispr Cloud as the default once you finish onboarding, with its cloud API handling transcription, sync, agent reasoning, settings and usage. For local AI dictation, switch both the speech model and the reasoning model to local ones.

Source: OpenWhispr, Network Allowlist, GitHub, OpenWhispr/openwhispr. Checked on 27 October 2026.

Weighing it against Voxmelt? The OpenWhispr and Voxmelt side-by-side goes through cost, licence, defaults, offline use and AI rewriting.

route 3 of 4

Build your own: whisper.cpp and Ollama

Free and as flexible as you like, if you are happy to write and look after the glue between them.

whisper.cpp runs OpenAI’s Whisper speech model in plain C/C++, under the MIT licence. Its README lists CPU-only inference, GPU support for NVIDIA, Vulkan and AMD ROCm, and Windows builds with MSVC and MinGW. Its whisper-stream example transcribes your microphone continuously; the README calls it a naive example.

Source: whisper.cpp, README, GitHub, ggml-org/whisper.cpp. Checked on 27 October 2026.

Ollama runs open language models on Windows, macOS and Linux, with a Windows installer, under the MIT licence. Its README shows a REST API on your own machine at localhost:11434, so a script can hand it a transcript with your clean-up instruction.

Sources: Ollama, README, GitHub, ollama/ollama; Ollama, MIT License, GitHub, ollama/ollama. Checked on 27 October 2026.

Its README also says it gives access to models locally or in the cloud, so check that the model you run is one that runs on your PC.

Source: Ollama, README, GitHub, ollama/ollama. Checked on 27 October 2026.

What is left is yours to build: a hotkey that starts and stops recording, a prompt for each kind of cleanup, and pasting the result where your cursor is. That is the part a finished app does for you.

route 4 of 4

Voxmelt: both steps in one Windows app

Voxmelt is our app, so weigh this section accordingly. It is built for this exact job: dictation on your CPU, then a local language model on your NVIDIA GPU that turns the transcript into finished text. The free routes may do some of this too; their pages are linked above.

Both steps on your PC

Speech is recognised on your CPU by default, with NVIDIA’s Parakeet v3 for English plus 24 European languages, or with Whisper for about 100 languages: tiny on the CPU, every size up to Large v3 Turbo on an NVIDIA GPU. The AI step runs through Ollama on your own PC, at 127.0.0.1:11434.

Templates instead of prompts

Pick a template and a tone instead of writing a prompt: 17 templates across 82 tones, for an email, a summary, meeting notes or a commit message, plus templates you write yourself. They work on what you dictate and on text you paste in.

The graphics card kept for the AI

With dictation on the CPU, your graphics card is left for the language model. If you run Whisper on the GPU as well, the app reads your card’s memory and decides on each run whether both models fit or one makes room.

Ollama, set up from the app

If Ollama is not installed, Voxmelt’s first-run setup can fetch its installer for you, if you choose. You pick the models from a catalog that includes Gemma, Qwen, Llama, Mistral, Phi and DeepSeek-R1.

Edit the result by voice

Say “make it shorter” to restyle what the AI wrote, or use the one-click tone shifts, such as shorter or sharper, instead of starting again.

What does leave the PC, listed

Your audio, transcripts, AI output and the prompts you write stay on it, and the app has no telemetry. What can leave: a hashed device fingerprint for the trial and licence check, your email and payment records if you buy, and anonymous update and model downloads, the Ollama installer among them if you choose it. A paid licence keeps working offline for 30 days after its last successful check.

The price, as the site shows it today

A 15-day trial of everything, then a free tier with 60 minutes of dictation and 10 AI runs a day on the starter models, with no card. Pro is $49 once, with unlimited AI runs on the everyday templates and the mid-size models; Studio is $99 once, with every model in the catalog. There is no subscription and nothing to cancel.

What it does not do: run the AI step on an AMD or Intel graphics card, or run on a Mac or Linux. It is not open source either, which OpenWhispr, whisper.cpp and Ollama are. Details are on the pricing page and the system requirements; what can leave the PC is listed on the privacy page.

questions

Local AI dictation on Windows, answered

What is local AI dictation?

Dictation where both models run on your own computer: a speech model turns your voice into text, and a language model cleans the text up or rewrites it. Cloud AI dictation runs one or both of those steps on a provider’s servers, so your audio or your text goes there.

Does Windows 11 have AI dictation built in?

On supported Copilot+ PCs, yes: fluid dictation runs small language models on the PC to correct grammar, add punctuation and remove filler words as you speak in English. Microsoft says availability varies by device, language, region, release status and installed Windows updates.

Sources: Use Voice Access to navigate Windows and applications, Microsoft Learn; Understand voice-powered productivity on Copilot+ PCs, Microsoft Learn. Checked on 27 October 2026.

Can AI dictation work offline?

Yes, when both models are on your PC: once they have downloaded, neither step needs a connection. With Voxmelt, dictation and the AI step both run with the Wi-Fi off. What does go online: model downloads, the trial and licence check (a paid licence keeps working offline for 30 days after its last successful check), your account and purchase if you buy, and update checks.

Do I need a graphics card for local AI dictation?

Not for the speech step: Parakeet runs on an ordinary CPU, and whisper.cpp supports CPU-only inference. The AI step is the heavy half. Voxmelt’s needs an NVIDIA GPU with 6 GB of video memory or more (Ampere or newer), and Windows’ fluid dictation needs a supported Copilot+ PC.

Sources: whisper.cpp, README, GitHub, ggml-org/whisper.cpp; Understand voice-powered productivity on Copilot+ PCs, Microsoft Learn. Checked on 27 October 2026.

What is Ollama, and do I need it?

Ollama is a free, MIT-licensed program that runs open language models on Windows, macOS and Linux and answers on your own machine at localhost:11434. Voxmelt uses it for the AI step and its setup can fetch the installer for you if you choose. For dictation alone you do not need it.

Sources: Ollama, README, GitHub, ollama/ollama; Ollama, MIT License, GitHub, ollama/ollama. Checked on 27 October 2026.

Will the AI change what I said?

It can. A language model writes new text: Voxmelt’s licence terms warn that AI post-processing may change meaning or add content you did not say, and Microsoft asks you to review what fluid dictation writes before you share it. Read the result before you send it. By default Voxmelt saves each session’s transcript and AI output on your PC, so you can compare the two.

Source: Understand voice-powered productivity on Copilot+ PCs, Microsoft Learn. Checked on 27 October 2026.

Is local AI cleanup as good as a cloud model?

We have not measured it, so we do not claim it. Our published benchmark covers the speech step only: 7.8% word error rate with Whisper large-v3 on a GPU, across 5 recorded clips, published on 16 July 2026. The cleanup depends on the model you run and on your text, so try it on your own words: Voxmelt’s free tier includes 10 AI runs a day.

Can I use an AMD or Intel graphics card?

Not for Voxmelt’s AI step, which needs an NVIDIA card; its dictation runs on the CPU, so that part needs no graphics card at all. OpenWhispr’s README says its local Whisper can use AMD and Intel GPUs through Vulkan, and whisper.cpp lists Vulkan and AMD ROCm support.

Sources: OpenWhispr, README, GitHub, OpenWhispr/openwhispr; whisper.cpp, README, GitHub, ggml-org/whisper.cpp. Checked on 27 October 2026.

Pick the route that fits your PC

On a supported Copilot+ PC, fluid dictation costs nothing extra, and OpenWhispr and the do-it-yourself route are free. If you want both steps in one Windows app with ready-made templates, Voxmelt’s free tier needs no card.

Direct installer, for PCs without the Store; Windows will warn because it is not code-signed.

Read how Voxmelt’s two engines share your hardware or dictation for developers on Windows.

sources

Where every outside fact comes from

Microsoft’s own pages and each project’s own repository, read on the date given. If one of them has changed since, tell us at help@voxmelt.com and we will fix this guide.

  1. Boost productivity with Voice Access and Fluid Dictation on Copilot+ PCs, Microsoft Learn. Checked on 27 October 2026.
  2. Understand voice-powered productivity on Copilot+ PCs, Microsoft Learn. Checked on 27 October 2026.
  3. Use Voice Access to navigate Windows and applications, Microsoft Learn. Checked on 27 October 2026.
  4. Windows 11 Insider Experimental Preview Build 26300.8493, Windows Insider Program, Microsoft Learn. Checked on 27 October 2026.
  5. OpenWhispr, README, GitHub, OpenWhispr/openwhispr. Checked on 27 October 2026.
  6. OpenWhispr, Network Allowlist, GitHub, OpenWhispr/openwhispr. Checked on 27 October 2026.
  7. whisper.cpp, README, GitHub, ggml-org/whisper.cpp. Checked on 27 October 2026.
  8. Ollama, README, GitHub, ollama/ollama. Checked on 27 October 2026.
  9. Ollama, MIT License, GitHub, ollama/ollama. Checked on 27 October 2026.

Voxmelt’s figures come from our benchmark, published 16 July 2026, the system requirements, the features page, the privacy page and the pricing page.