Moonshine Streaming Tiny Arabic — Tadabbur ONNX

Quantized ONNX conversion of moonshine-ai/moonshine-streaming-tiny-ar, pinned to revision 2d3a986306561346ef0bf0018d23238944faf087.

This is a runtime-optimized conversion, not a Qur'an fine-tune. Encoder sliding-window attention is preserved. MatMul/Gemm and embedding Gather weights are quantized; convolutions remain floating point.

Runtime contract

Use ONNX Runtime directly (these files are not a Transformers.js pipeline export):

• encoder_q8.onnx: input_values float32 [1, samples] → hidden_states float32 [1, frames, 320].
• decoder_q8.onnx: input_ids int64 [1, tokens], encoder_hidden_states float32 [1, frames, 320] → final-position logits float32 [1, 12288].
• Audio: mono 16 kHz, pad to a multiple of 80 samples. Batch size one; all supplied frames valid.
• Start with token 1, append argmax logits, stop at token 2. Limit generation to floor(duration_seconds * 6.5) + 2 tokens.
• Decode with tokenizer.json, skipping special tokens.

Designed for short recitation phrases. The Tadabbur integration limits phrases to seven seconds and re-decodes growing phrases; this export does not expose stateful streaming or decoder KV caches.

Validation

FP32 ONNX was compared with PyTorch, with maximum logit difference 0.00001574 on shape checks. A single six-second Alafasy 1:1 clip produced بسم الله الرحمن الرحيم with both fp32 and q8. The final browser smoke test took 365 ms on the development machine. This is not a representative accuracy or latency benchmark; no improvement over Qur'an-tuned Whisper is established.

Reproduction scripts and detailed limitations are in conversion/. The precise source revision and export metadata are in manifest.json.

Attribution

Original model and tokenizer: Moonshine AI, MIT license. This repository provides a converted artifact of those weights, not a newly trained model.

License and attribution

Original model and tokenizer: Moonshine AI (Useful Sensors, Inc.), MIT license. See the LICENSE file in this repository for the copyright notice and full license text.

• Base model: moonshine-ai/moonshine-streaming-tiny-ar, revision 2d3a986306561346ef0bf0018d23238944faf087
• Changes made: ONNX export and quantization for browser inference. No retraining or fine-tuning.