benderrodriguez/whisper-large-v3-turbo-webgpu

whisper-large-v3-turbo, WebGPU f16 weights

ivrit-ai/whisper-large-v3-turbo repacked for whisper-gpu, a from-scratch TypeScript + WGSL Whisper engine that runs in the browser. These files are not loadable by transformers or whisper.cpp — they are a flat tensor dump laid out for WebGPU storage buffers.

Layout

file	contents
manifest.json	model config, audio config, shard table, and a tensor index (name → shard, offset, shape, dtype)
tokenizer.json	vocabulary, merges, special tokens, language ids
weights-*.bin	raw tensor bytes, packed f16

Weights are stored as packed f16 and read in the shaders via unpack2x16float with f32 accumulation, so no adapter needs the optional shader-f16 feature. Every kernel fits the WebGPU default limits (256 invocations per workgroup, 16 KiB workgroup storage, 128 MiB storage binding), which is what lets the same shaders run on iOS Safari.

Use

import { Whisper } from "whisper-gpu";
import { FetchSource } from "whisper-gpu/loader";

const src = new FetchSource(
  "https://huggingface.co/benderrodriguez/whisper-large-v3-turbo-webgpu/resolve/main",
);
const whisper = await Whisper.load(ctx, src);
for await (const seg of whisper.transcribe(audio, { language: "he" })) console.log(seg.text);

Licence

Apache 2.0, inherited from the base model.