Basira — whisper-base-ar-quran ONNX (fp16 / q4 / int8)

ONNX exports of tarteel-ai/whisper-base-ar-quran (whisper-base fine-tuned on tarteel-ai/everyayah) for the on-device Live Tartil engine at basira.institute. Runs fully in the browser via transformers.js — recitation audio never leaves the device.

Files

Variant	Encoder	Decoder	Backend
*_fp16	41 MB	158 MB	WebGPU on adapters with shader-f16 (modern phones, RTX/Apple-Silicon class)
*_q4	19 MB	140 MB	WebGPU on adapters without shader-f16 (e.g. Pascal-class NVIDIA) — MatMulNBits, block_size 32, symmetric
*_quantized	23 MB	79 MB	WASM (dynamic int8)

The fp16 and int8 files are mirrored from aaqibhabib/whisper-base-ar-quran-onnx; the q4 files are quantized from that repo's fp32 export with onnxruntime's MatMulNBitsQuantizer.

Usage (transformers.js)

import { pipeline } from "@huggingface/transformers";
const asr = await pipeline(
  "automatic-speech-recognition",
  "iqbalaesthetic/Basira",
  { device: "webgpu", dtype: "fp16" }, // or dtype: "q4"; wasm wants "q8"
);

Note: generation_config.json ships without alignment_heads; they are an architecture constant of whisper-base and are patched in at runtime by the Basira worker (see WHISPER_BASE_ALIGNMENT_HEADS in the app source).