whisper-large-v3-turbo-lb (ONNX, for the browser)
ONNX weights of SitrakaForler/whisper-large-v3-turbo-lb, for transformers.js on WebGPU. Same model, same accuracy figures (see the base model card); only the format and the weight precision change.
| file | precision | size |
|---|---|---|
onnx/encoder_model_q4.onnx |
4-bit MatMul weights, block 32 | 425 MB |
onnx/encoder_model_fp16.onnx |
16-bit | 1.27 GB |
onnx/decoder_model_merged_q4.onnx |
4-bit MatMul weights (If branches stay fp32) | 334 MB |
import { pipeline } from "@huggingface/transformers";
const asr = await pipeline("automatic-speech-recognition", "SitrakaForler/whisper-large-v3-turbo-lb-onnx", {
device: "webgpu",
dtype: { encoder_model: "q4", decoder_model_merged: "q4" }, // or encoder_model: "fp16"
});
const { text } = await asr(audio16kFloat32, { language: "luxembourgish", task: "transcribe", chunk_length_s: 30 });
The 4-bit files have not been scored on FLEURS separately. On a 2'26" radio report they gave a transcript close to the CTranslate2 int8 export of the same model.
Demo that runs entirely in the browser: SitrakaForler/lb-whisper-demo.
Export script: finetune/export_onnx.py in github.com/Sitraka17/lb-whisper. MIT.