Valnivo-labs/whisper-small-lb-onnx

Whisper small, Luxembourgish — ONNX, 8-bit

An ONNX export of unilux/whisper-small-v1-luxembourgish (commit 33f11e93ab16), developed by the LuxASR team at the University of Luxembourg (Department of Humanities). Nothing about the model was changed except its format: exported with Optimum, the encoder's weights quantised to uint8 and the merged decoder's to int8, per tensor, plus the encoder in fp16 for WebGPU — the layout Transformers.js loads as dtype: 'q8' or dtype: { encoder_model: 'fp16', decoder_model_merged: 'q8' }.

It exists for the optional speech check in Moien, a free Luxembourgish course, which runs it entirely on the learner's own device: recordings are never uploaded. Built reproducibly with Optimum 1.17.1, Transformers 4.38.2, PyTorch 2.5.1 and ONNX Runtime 1.30.0.

Licence

OpenMDW 1.0, the licence of the original model: the full text is in LICENSE, and the original model card is kept unchanged as UPSTREAM-README.md. All credit for the model belongs to its authors.

Usage

import { pipeline } from '@huggingface/transformers'
const asr = await pipeline('automatic-speech-recognition', 'Valnivo-labs/whisper-small-lb-onnx', { dtype: 'q8' })
const { text } = await asr(samples16kHz, { language: 'luxembourgish', task: 'transcribe' })