Almost certainly yes. Pick your system below to see which compute backend WhisPaste uses, how much graphics memory it wants, and what happens when the GPU is missing, old or simply broken.
Short answer: WhisPaste never requires a GPU. On Windows and Linux it accelerates through Vulkan, which covers NVIDIA, AMD and Intel with a single build; on Apple Silicon it uses Metal. Everywhere else — outdated driver, no dedicated card, a machine that simply refuses — it runs the model on the CPU. 8 GB of RAM is the only hard requirement, and it is checked at startup.
JavaScript is off, so all three possible outcomes are shown below. Pick the one that matches your system.
WhisPaste ships a Vulkan build of whisper.cpp for Windows and Linux. Vulkan is one code path for NVIDIA, AMD and Intel alike, so a non-NVIDIA card is not a second-class citizen here: if your driver exposes Vulkan, the model runs on the GPU. Budget roughly 900 MB of graphics memory for the compact model, about 1.5 GB for balanced and about 2.6 GB for premium.
The Mac build targets Apple Silicon and accelerates through Metal. Apple Silicon shares one pool of memory between processor and graphics, so there is no separate video-memory budget to plan around: an 8 GB Mac already carries the balanced model comfortably.
Without a usable GPU driver WhisPaste runs the model on the processor. That is a supported configuration, not a failure state — the compact model exists for exactly this case, and the requirement stays at 8 GB of RAM and a 64-bit processor. If speed matters more to you than language coverage, the second on-device engine, Parakeet, is CPU-only by design and several times faster than Whisper on the same machine, covering 25 European languages instead of 99.
Older GeForce cards are covered too. The 400 to 700 series (Fermi and Kepler) predate the CUDA generation most prebuilt speech binaries require, which is one reason WhisPaste ships Vulkan instead: those cards keep working rather than crashing at launch.
AMD cards run the exact same Vulkan build as everything else — no vendor-specific download, no extra runtime to install first.
Intel integrated graphics count as a GPU here: if the driver exposes Vulkan, the model runs on the iGPU. Because integrated chips borrow system memory rather than owning dedicated video memory, WhisPaste stays deliberately conservative about pushing the largest model onto them.
Detection happens once at startup, and the choice is made without asking you: Vulkan on Windows and Linux, Metal on Apple Silicon, processor whenever no compatible GPU is found or detection has not finished yet. You can override it — the “CPU only” setting forces the processor path on any machine, which is the fastest way to rule the GPU out when you are debugging something else.
Two safety nets keep a bad driver from taking the app down with it. If loading a model hangs on the GPU — a cold-start shader compile that never finishes, for example — WhisPaste retries once on a dedicated processor-only engine. And if loading crashes the GPU twice in a row, acceleration is switched off permanently on that machine: the app keeps transcribing on the processor instead of trapping you in a crash loop. Nobody has to read a log file for either of those to happen.
The floor is 8 GB of RAM and a 64-bit processor on Windows 10, macOS or Ubuntu Linux; the app checks memory at startup and tells you plainly if the machine is below the line. Everything above that is optional. Graphics memory only matters if you want the larger models, and it scales with the model rather than with the app.
| Tier | Graphics memory |
|---|---|
| Compact (whisper-small) | ~900 MB |
| Balanced (whisper-medium) | ~1.5 GB |
| Premium (whisper-large-v3-turbo) | ~2.6 GB |