What is WebGPU, and why your browser can now run a speech model

WebGPU is a browser feature that lets a web page use your computer’s graphics card for general calculation, not just for drawing pictures. That one change is why a speech recognition model can now run inside a browser tab at a usable speed, with nothing uploaded. It ships on by default in Chrome and Edge on Windows, macOS and ChromeOS, in Safari 26, and in Firefox on Windows. This post explains what a graphics card does, how WebGPU differs from the older WebGL, which browsers have it, what can stop it working on your machine, and how to check.

A graphics card is a very wide calculator

Your computer has two chips that can do arithmetic. The processor, the CPU, has a handful of cores, and each one is fast and flexible. It runs your browser, your spreadsheet and the operating system, mostly one instruction after another.

The graphics card, the GPU, is built the other way round. It has thousands of small, simple units that all do the same operation on different numbers at the same time. That is exactly what drawing a screen needs, because every pixel gets roughly the same treatment. It is also what a neural network needs, because most of a model’s work is multiplying large grids of numbers together, and every cell in the result can be worked out in parallel.

For this kind of maths the card beats the CPU by a wide margin. Until recently a web page had no good way to ask for it.

WebGPU gives a page the whole card, not just the drawing part

Web pages have been able to use the graphics card since WebGL arrived, but WebGL was designed for drawing. MDN’s description of the situation is blunt: “WebGL does not handle general-purpose GPU computations very well”, and because it is based on OpenGL, “there are no more updates planned to OpenGL (and therefore WebGL), so it won’t get any of these new features.”

WebGPU is the successor. MDN defines it as an API that lets developers “use the underlying system’s GPU to carry out high-performance computations and draw complex images”. It is modelled on the modern native graphics interfaces, Direct3D 12 on Windows, Metal on Apple hardware and Vulkan elsewhere, and it adds something WebGL never had: a compute pipeline. Apple’s WebKit team put it plainly when Safari 26 shipped: WebGPU “adds compute shaders, which allow general purpose computations on the GPU, something not previously possible with WebGL.”

In practice a page asks the browser for an adapter, which stands for a physical graphics card and its driver, then asks that adapter for a device it can send work to. If either step comes back empty, WebGPU is not available on that machine.

Why a speech model wants the GPU

A speech recognition model like OpenAI’s open-source Whisper is a neural network. It takes a slice of audio, turns it into a grid of numbers, and pushes that grid through many layers of multiplication to arrive at the most likely words. On a CPU that arithmetic happens a few numbers at a time. On a GPU it happens thousands of numbers at a time.

Chrome’s launch post for WebGPU claimed “more than three times improvements in machine learning model inferences”, though it does not say against what baseline. In our test, a desktop with a graphics card transcribed 71 seconds of audio in 41 to 50 seconds including loading the model, about 1.5 times real time. A thin laptop with only integrated graphics takes roughly as long as the recording. Without WebGPU the same model would be too slow to be worth offering.

Because the arithmetic runs on your card, the site needs no server to do it and nothing is uploaded. That is what makes a free transcriber with no time cap possible at all.

Which browsers ship WebGPU, and since when

The WebGPU working group keeps a status page, last updated 13 August 2026 when we checked, and the browser makers publish release notes. This is where things stand.

One footnote: Chrome’s announcement said the Windows release covers “Windows devices with Direct3D 12 support”, so a very old Windows install is left out even with a current Chrome. And shipping in a browser is not the same as working on your machine.

What stops WebGPU working on a particular machine

Chrome’s troubleshooting page lists the reasons a page gets no adapter back. The ones that affect ordinary users are:

• Hardware acceleration is switched off. Chrome disables WebGPU when “Use graphics acceleration when available” is turned off in chrome://settings/system.
• The graphics card is on the blocklist. chrome://gpu will say the GPU hardware has been specifically blocklisted, or that “WebGPU has been disabled via blocklist or the command line”.
• Old drivers. If chrome://gpu reports “WebGPU: Software only, hardware acceleration unavailable”, Chrome’s advice is to update your GPU drivers.
• The platform is not supported yet. Windows on ARM and most Linux setups fall here unless you turn on experimental flags.
• The GPU process has crashed several times. Reloading the page or restarting Chrome fixes this one.

Virtual machines and remote-desktop sessions tend to land in the same bucket, because the browser is usually handed a virtual display device rather than the real card. Chrome’s page does not call them out by name, so treat that as a rule of thumb rather than a documented rule.

The Web3D Survey collects reports from visitors to 23 contributing sites and records whether a working WebGPU adapter came back. When we checked, its overall figure was 82.73%. By operating system it was 87.5% on Windows, 91.1% on macOS and only 17.1% on Linux. By browser it was 92.0% on Edge, 84.1% on Chrome, 81.5% on Safari and 60.4% on Firefox. The caveat: the contributing sites are 3D and graphics projects, so the audience skews toward machines with capable graphics, and the survey does not publish a sample size or time window. Treat the numbers as an upper bound for a general audience.

How to check your own machine

In Chrome, type chrome://gpu into the address bar and find the WebGPU line. Chrome’s documentation says to “make sure you can read ‘WebGPU: Hardware accelerated’”. If it says “Software only, hardware acceleration unavailable”, update your graphics driver. If it says disabled via blocklist, your card or driver is on Chrome’s list. Chrome documents a flag, chrome://flags/#enable-unsafe-webgpu, that overrides the blocklist, but the name tells you how supported that path is.

What this means for the transcriber

FreeTranscribe runs OpenAI’s open-source Whisper model, the base English version, on your graphics card through WebGPU. Everything above sets its limits.

It works today on desktop Chrome or Edge with a working adapter. Safari 26 and Firefox on Windows ship WebGPU, but we have not added those code paths yet, and phones are not supported. It is English only. The model is about 200 MB, downloaded once and cached by the browser, and it is the base model, so names, technical terms, heavy accents and noisy rooms are its weak spots. Read the transcript through before relying on it.

In return there is no upload, no account and no time or length limit, and speed depends on your hardware. If your machine passes the chrome://gpu check above, drop a file on the tool and see how long it takes.

Frequently asked questions

Is WebGPU the same as WebGL? No. WebGL is the older API and is built for drawing. WebGPU is its successor, is modelled on Direct3D 12, Metal and Vulkan, and adds compute shaders for general calculation. MDN notes there are no further updates planned for OpenGL, and therefore WebGL.

Do I need a gaming graphics card? No. Any card the browser can use for WebGPU will do, including a laptop’s integrated graphics. A discrete card is faster: in our test a desktop card ran at about 1.5 times real time, a thin laptop at about the length of the recording.

Why does the tool say WebGPU is unavailable when I have Chrome? The usual causes are hardware acceleration turned off in settings, a blocklisted or out-of-date driver, a virtual machine or remote-desktop session, or Linux, where Chrome only enables WebGPU on some hardware. Open chrome://gpu and read the WebGPU line.

Does WebGPU send anything to the graphics card maker or to Google? WebGPU is a way for a page to use the card in your own computer. The transcriber’s processing happens in the tab, and you can watch the browser’s network panel stay empty while it runs.

Sources, checked 14 September 2026

• https://github.com/gpuweb/gpuweb/wiki/Implementation-Status : WebGPU working group shipping status per browser and platform, last updated 13 August 2026
• https://developer.mozilla.org/en-US/docs/Web/API/WebGPU_API : MDN definition of WebGPU, comparison with WebGL, adapter and device concepts
• https://developer.chrome.com/blog/webgpu-release : Chrome 113 release post, platforms, Direct3D 12 requirement, machine learning claim
• https://developer.chrome.com/blog/new-in-webgpu-113 : Chrome 113 shipping on ChromeOS, macOS and Windows
• https://developer.chrome.com/docs/web-platform/webgpu/troubleshooting-tips : Chrome’s reasons for a null adapter, the chrome://gpu check and the unsafe-webgpu flag
• https://webkit.org/blog/17333/webkit-features-in-safari-26-0/ : WebKit post on Safari 26, WebGPU on macOS, iOS, iPadOS and visionOS
• https://www.firefox.com/en-US/firefox/141.0/releasenotes/ : Firefox 141 release notes, WebGPU enabled on Windows
• https://www.firefox.com/en-US/firefox/145.0/releasenotes/ : Firefox 145 release notes, WebGPU on macOS 26 on Apple Silicon
• https://web3dsurvey.com/webgpu : Web3D Survey working-adapter percentages, as recorded in our research notes on 14 September 2026; the site itself was unreachable from our network at the time of writing
• Speed and accuracy figures are from our own test of the tool, described in the text