whisper-small-setswana-onnx

whisper-small-setswana (ONNX for Transformers.js)

ONNX export (q8 / int8 quantized) of misterkissi/whisper-small-setswana, packaged for in-browser speech recognition with 🤗 Transformers.js.

This lets a Setswana (Tswana) Whisper model run entirely client-side in the browser (WebGPU with WASM fallback) — no server, no audio leaving the device.

Files

• onnx/encoder_model_quantized.onnx (~88 MB)
• onnx/decoder_model_merged_quantized.onnx (~186 MB)
• tokenizer / preprocessor / config JSON

Usage

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

const asr = await pipeline(
  'automatic-speech-recognition',
  'Hydramus/whisper-small-setswana-onnx',
  { device: 'webgpu', dtype: 'q8' }
);
const { text } = await asr(float32AudioAt16kHz);

Attribution & license

• Base model: misterkissi/whisper-small-setswana, a fine-tune of openai/whisper-small.
• License: CC-BY-NC-SA-4.0 — same as the base model (share-alike). Attribution required; non-commercial use only.
• This repository only re-packages the original weights as quantized ONNX; all model credit belongs to the original author.

Quality note

The base fine-tune reports WER ≈ 0.48 with signs of overfitting, so transcripts are a useful rough draft for Setswana but not production-clean. Quantization to int8 preserves the original behaviour (verified: outputs match the PyTorch model up to minor rounding).