ONNX export of Shakkala (version 3) by Ahmad Barqawi, MIT-licensed. Converted and verified by TigreGotico: the ONNX is 100 % output-identical to the original Keras model on held-out sentences.

I/O. Input input — float32, shape [batch, 315], character ids post-padded with 0. Output — [batch, 315, 28] class logits. Encode with input_vocab_to_int.json (<UNK> fallback), pad to 315; decode argmax via output_int_to_vocab.json (skip <PAD>), then interleave each diacritic onto its character.

Architecture. Embedding(288) → 3× BiLSTM (288/144/96 per direction) + BatchNorm → Dense(28). ~2.5 M params, fixed 315-length.

Benchmark. ~5.65 % DER on the broad TigreGotico/arabic_diacritized_text test split.

Part of the Arabic Diacritizers / Tashkeel collection. Original work © Ahmad Barqawi (MIT); this is a faithful re-export with attribution.