whisper-tiny-ar-quran-onnx

Transformers.js-compatible ONNX conversion of tarteel-ai/whisper-tiny-ar-quran. The decoder exports cross-attention tensors and the generation config contains Whisper alignment heads, enabling word-level timestamps in Transformers.js.

Usage

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

const transcriber = await pipeline(
  "automatic-speech-recognition",
  "Sharjeelbaig/whisper-tiny-ar-quran-onnx",
  { dtype: "q8" },
);

const result = await transcriber(audioFloat32Array, {
  language: "ar",
  task: "transcribe",
  return_timestamps: "word",
});

The repository contains fp32 and dynamically quantized q8 encoder/merged decoder graphs. q8 is recommended for local CPU inference.

Validation

The q8 artifact was loaded with @huggingface/transformers 4.2.0 in Node.js. It returned the correct text and four timestamped words for Qur'an 1:1. A second-reciter test on Husary's complete Sūrah 112 found all four verses and aligned all 15 canonical words from 16 recognized tokens. The artifact was also exercised end to end by transcribe-quran, including offline cached inference, canonical matching, ASS generation, and FFmpeg/libass video rendering.

SHA-256:

Conversion

convert-model.py is the reproducible development-only exporter. It enables eager attention and explicitly exposes encoder, decoder, and cross-attention outputs before Optimum merges the cached and non-cached decoder graphs.

uv run --python 3.12 \
  --with torch --with accelerate --with 'optimum-onnx[onnxruntime]' \
  python convert-model.py --output export

uvx --python 3.12 --with torch --with accelerate \
  --from 'optimum-onnx[onnxruntime]' optimum-cli onnxruntime quantize \
  --onnx_model export --arm64 --per_channel -o quantized

The published onnx/*_quantized.onnx files are the encoder and merged decoder from that quantized output, using Transformers.js's _quantized q8 filename convention. The full generation metadata comes from the matching multilingual Whisper Tiny generation config.

The source checkpoint is Apache-2.0. See its model card for reported training and evaluation details and limitations.