Offline Speech to Text on Windows: Running Whisper Locally
For dictation that never touches the internet, Whisper running locally on your own machine is the most practical option available today. Here's how the model sizes compare and how to set it up.
The VivixCore team · September 19, 2026 · 5 min read
Sending your voice to a server is a reasonable thing to be uneasy about, especially if you're dictating client details, financial information or anything else you'd rather not have leave your machine. The good news is that offline speech to text on Windows has gotten genuinely practical, not just theoretically possible.
Cloud dictation vs. running everything locally
Cloud based dictation, where audio is sent to a transcription service and text comes back, has two clear advantages: it needs no setup and no particular hardware, and it can use larger, more frequently updated models than most personal computers can run. The trade off is that your audio does leave your machine to be processed.
Running everything locally flips that trade off. Nothing leaves your computer, which matters if you're dictating anything sensitive, but you're now relying on your own hardware, and setup takes a bit more effort up front.
Neither option is universally right. It depends on what you're dictating and what your machine can handle.
There's also a middle consideration worth naming: latency and reliability. Local processing doesn't depend on your internet connection at all, which matters if you're dictating on a flight, in a location with unreliable Wi-Fi, or somewhere a client's network policy blocks outbound traffic to third party services entirely. Cloud transcription needs a connection, but in exchange it needs nothing from your own hardware.
What Whisper actually is
Whisper is an open source speech recognition model, originally released by OpenAI, that transcribes spoken audio into text. Because it's open source, it can run entirely on your own hardware rather than through an API, which is what makes fully offline speech to text realistic on a normal Windows PC.
Whisper comes in several sizes, each one trading accuracy for speed and hardware requirements. The right size depends entirely on what machine you're running it on.
Try it while you read
Hold a key, talk, and watch clean text land in any Windows app.
Choosing a model size for your hardware
| Model | Parameters | Approximate VRAM | Relative speed |
|---|---|---|---|
| Tiny | 39M | ~1 GB | Fastest |
| Base | 74M | ~1 GB | Very fast |
| Small | 244M | ~2 GB | Fast |
| Medium | 769M | ~5 GB | Moderate |
| Large | 1.55B | ~10 GB | Slowest, most accurate |
| Turbo | 809M | ~6 GB | Fast, near large level accuracy |
A rough way to think about it: tiny and base are for quick notes on modest hardware, where speed matters more than catching every word perfectly. Small and medium are a reasonable middle ground for most laptops with a dedicated GPU. Large gives the best accuracy but needs meaningfully more VRAM and runs slower. Turbo, a newer, optimized variant, gets close to large's accuracy at a fraction of the resource cost, which makes it a strong default for anyone unsure where to start.
It's worth testing more than one size on your own voice and typical background noise before settling on a default. Accuracy differences between model sizes tend to show up most with accents, background noise or fast speech, and less with slow, clear dictation in a quiet room, so the "right" size for you depends partly on how and where you actually talk.
Local Whisper for speech, a local LLM for the rewrite
Speech to text is only half of what makes dictation useful day to day. The other half, turning a raw transcript into clean, on-tone text, normally runs through an AI rewrite step. Keeping that step offline as well means pairing local Whisper with a local language model, run through something like Ollama or LM Studio, both of which let you download and run open models directly on your machine.
The two pieces work together: Whisper turns your voice into a raw transcript, and the local model handles the same kind of cleanup and tone rewriting that a cloud based rewrite step would do, filler words removed, punctuation added, tone matched to the app you're typing into. Neither the audio nor the rewritten text ever needs to leave your computer.
This is exactly the setup VivixCore supports: local Whisper for transcription, Ollama or LM Studio for the rewrite, both running entirely on your own hardware. See the full breakdown under your AI, your choice on the homepage.
Setting it up in VivixCore
Getting a fully offline setup running takes a few steps:
- Install Ollama or LM Studio and download a model that fits your hardware, following either tool's own setup guide.
- Download a Whisper model in VivixCore's settings, choosing a size based on the table above.
- Point VivixCore at your local model instead of VivixCore Cloud, in the AI engine settings.
- Test with a normal sentence before relying on it for anything important, since local model quality varies more by hardware than cloud services do.
Once it's set up, dictation works exactly the same way from your end: hold a key, speak, get clean text. The only difference is where the processing happens.
Expect the initial setup, downloading a model through Ollama or LM Studio and a Whisper model of your chosen size, to take longer than the day to day experience afterward. Model files range from a few hundred megabytes for the smallest options to several gigabytes for the larger ones, so it's worth doing this step on a stable connection before you need to rely on the offline setup for something time sensitive.
When your own API key is the better middle ground
Fully local setups aren't the only privacy conscious option. Connecting your own API key, from a provider like Groq, OpenAI, Gemini, Anthropic, DeepSeek, Cerebras or OpenRouter, keeps your usage tied to an account you control and your key stored only on your own device, without needing local hardware capable of running large models well.
This tends to suit people who want more control than a shared cloud allowance but don't have a machine with enough VRAM to run larger Whisper or language models comfortably. It's a genuine middle ground, not a compromise: rewrite quality with your own key can match or exceed a smaller local model, while still keeping your credentials entirely in your own hands.
It's also worth remembering that these approaches aren't mutually exclusive over time. Plenty of people start on VivixCore Cloud to get moving quickly, move to their own API key once they know how much they dictate and want more control over cost and provider, and only go fully local later if privacy requirements change or a specific project calls for it. Nothing about switching between them requires starting over, since the dictation workflow itself stays identical the whole way through.
Whichever route fits your setup, whether that's VivixCore Cloud for zero configuration, your own API key for control without local hardware demands, or fully offline with local Whisper and Ollama for nothing leaving your machine, the dictation experience itself stays the same. Only where the processing happens changes.
For the basics of getting dictation running on Windows in the first place, see our guide on voice typing on Windows 11.
Frequently asked questions
Can speech to text work completely offline on Windows?+
Yes. Running Whisper locally handles the speech to text step without an internet connection, and pairing it with a local model through Ollama or LM Studio keeps the rewrite step offline too.
Which Whisper model size should I use?+
It depends on your hardware. Smaller models like base or small run comfortably on a laptop with modest specs, while medium and large need more VRAM but produce noticeably more accurate transcripts, especially for accents or background noise.
Is local Whisper as accurate as cloud transcription?+
The larger local models get close, particularly the large and turbo variants, though cloud transcription services can use larger models than most personal machines can run comfortably and update them more often.
Do I need a powerful GPU to run Whisper locally?+
Not for the smaller models. Tiny and base run fine on modest hardware. Medium and large benefit from a dedicated GPU with several gigabytes of VRAM, but they're not out of reach for a typical gaming laptop.