Runs in the browser · Speech to text

Run Whisper Small in your browser

Whisper Small downloads 462.7 MB over WebGPU (fp16), or 237.5 MB over WebAssembly (uint8), to transcribe speech to text entirely in the tab with Transformers.js. No install, no server.

Speech to text · onnx-community/whisper-small

WebGPU download

462.7 MB

fp16

WASM download

237.5 MB

uint8

Parameters: 241.7M
Pipeline task: automatic-speech-recognition

Reading

The WASM build downloads smaller here: 237.5 MB (uint8) against 462.7 MB (fp16) for WebGPU. The CPU-fallback quant compresses tighter than the GPU pick for this model.

The quant ladder spans 3.9x: 237.5 MB (uint8) to 923.3 MB (fp32).

Live check this browser

Checking for WebGPU support…

All measured variants

Quant Download size

Variant	Size
uint8 WASM pick	237.5 MB
bnb4	273.7 MB
q4	285.5 MB
fp16 WebGPU pick	462.7 MB
q8	475.1 MB
fp32	923.3 MB

Sizes measured from the HuggingFace API file tree for onnx-community/whisper-small, not estimated.

Use it with Transformers.js

import { pipeline } from "@huggingface/transformers";

const pipe = await pipeline("automatic-speech-recognition", "onnx-community/whisper-small", {
  device: "webgpu", // usually falls back to "wasm"; wrap in try/catch for production
  dtype: "fp16", // use "uint8" for the WASM build
});

const result = await pipe(/* your input */);

Requires npm install @huggingface/transformers (or the CDN build).

Sources

Same job, different size

← Whisper Base 139.3 MB Whisper Large v3 Turbo → 537.4 MB

FAQ

WebGPU or WASM for Whisper Small, in practice?

Chrome, Edge, and Safari 26+ run the WebGPU build (fp16) on the GPU. Firefox has WebGPU on Windows and Apple Silicon Macs but not everywhere yet. Any browser without a WebGPU adapter falls back automatically to the WebAssembly build (uint8) on CPU: same model, slower to load, slower to run.

What does fp16 mean for Whisper Small?

The WebGPU build here uses fp16: 16-bit floating point (half precision): smaller than fp32 with effectively no quality loss. The WASM fallback uses uint8: 8-bit integers: about a quarter the size of fp32, with a smaller quality trade than 4-bit.

Sizes measured 2026-08-01 from the HuggingFace API. Last validated 2026-08-03. See all browser models or the methodology.