whisper-large-v3-turbo-webgpu

Whisper large-v3-turbo · ONNX for browser WebGPU

ONNX export of openai/whisper-large-v3-turbo in the Transformers.js repo layout, prepared for in-browser WebGPU inference.

Files under onnx/:

• encoder_model_fp16.onnx — encoder in float16 (1.27 GB)
• decoder_model_merged_q4.onnx — merged decoder, 4-bit weight quantization (MatMulNBits, block size 32; 375 MB)
• decoder_model_merged_fp16.onnx — merged decoder in float16 (477 MB), alternative for runtimes where the q4 kernels misbehave

Motivation: on some GPU/driver generations (observed on recent Apple Silicon with Chrome's Metal WebGPU backend) the widely used community ONNX exports of Whisper produce corrupted output on WebGPU while the same files run correctly on WASM. This export, produced with a current Optimum toolchain (fp32 export → fp16 conversion with subgraph-aware Cast fixing → MatMulNBits quantization), runs correctly on WebGPU on those machines.

Usage with Transformers.js:

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

const asr = await pipeline("automatic-speech-recognition",
  "striimit/whisper-large-v3-turbo-webgpu", {
    device: "webgpu",
    dtype: { encoder_model: "fp16", decoder_model_merged: "q4" },
  });
const { text } = await asr(float32Audio16kHz, { task: "transcribe", language: "fi" });

License: MIT, inherited from the original OpenAI Whisper model.