### NOTICE

**Ghost Assistant: Optimized OmniASR v2 Pipeline**

**Original Model:** Weights and architecture by **Meta Research**.

**Quantization Engine:** This project utilizes the `MatMulNBitsQuantizer` and infrastructure provided by **ONNX Runtime** (Microsoft).

**Technical Modifications:**
This distribution contains significant independent engineering and optimization performed by **Edison dos Santos**:

* **Standalone ONNX Architecture:** Manual graph reconstruction and decoupling from original research libraries, allowing for direct usage with standard ONNX Runtimes.
* **LayerNorm Compatibility Patch:** Manual epsilon adjustment (1e-5) within the model graph to ensure numerical stability on mobile NPUs and generic hardware.
* **4-bit Quantization (Q4):** Specialized Weight-Only Quantization (WOQ) using a `block_size=32` strategy, specifically tuned for ARM-based architectures (optimized for Dimensity 6300).
* **Vocabulary Extraction:** Independent SentencePiece-to-JSON mapping for standalone CTC decoding without external dependencies.

**Copyright (c) 2026 Edison dos Santos**