WebGPU vs WebAssembly for Whisper
Both run Whisper on your device. The choice changes which hardware does the work, browser compatibility and how you recover from failures.
Start with WebAssembly, the default backend. Try WebGPU on supported hardware if you want to compare processing speed. Switch back if model loading or transcription fails.
CPU or GPU: the practical differences
| Question | WebAssembly | WebGPU |
|---|---|---|
| Primary compute | CPU through the browser runtime | Supported GPU through the browser |
| Compatibility | Broader fallback | Depends on browser, OS, driver and hardware support |
| Setup in Whisper Web | Default Tiny Q8 path | Optional backend selected by the user |
| Best first choice | Unknown or constrained device | Supported desktop device that needs more throughput |
| Failure recovery | Try a smaller model or another supported file | Return to WebAssembly if initialization fails |
Model size still matters
Tiny, Base and Small models differ in download size, memory use and processing time. WebGPU may improve throughput on supported hardware, but it does not make every model suitable for every device.
The first run downloads the selected model. Whisper Web can reuse the browser cache later, unless the browser clears it or storage is unavailable.
Recover from a failed run
If WebGPU is unavailable
Use WebAssembly. The transcription task does not require WebGPU.
If the device runs out of memory
Choose a smaller model and close other heavy tabs before retrying.
If the first run appears slow
Separate model-download time from transcription time before comparing backends.
Common questions
Is WebGPU required for Whisper Web?
No. WebAssembly is available as the compatible local fallback.
Is WebGPU always faster?
No. Performance depends on the browser, GPU, driver, model and recording. Compare on the device you actually plan to use.
Do both backends keep audio local?
Yes. In Whisper Web local mode, both backends process the selected media on the current device.