| # Third-party components | |
| The source modules `omini/models/vision_transformer.py` and | |
| `omini/models/spatial_ast.py` originate from the SpatialAST components in | |
| SLAM-LLM. The latter includes adapted buffer-based spectral operators. | |
| Upstream source: https://github.com/X-LANCE/SLAM-LLM/tree/a68e78f93a12a678438c8b43974fd987abb66927/src/slam_llm/models/SpatialAST | |
| The upstream MIT license is reproduced in `licenses/SLAM-LLM-MIT.txt`. | |
| Existing public third-party attribution is retained; it is not a statement of | |
| submission authorship. | |
| The binaural frontend follows the BAT / Spatial-AST feature construction. | |
| The spectral shims follow torchlibrosa's Fourier and mel-filter mathematics; | |
| librosa supplies the mel filters. timm supplies transformer utility layers. | |
| Hugging Face Transformers provides the native visual backbones and processors. | |
| These installed dependencies and any externally obtained model weights retain | |
| their respective licenses. This archive does not redistribute model weights. | |
| The `omini` namespace is retained solely for source import compatibility. Only | |
| the encoder dependency closure and canonical waveform preprocessing needed by | |
| this package are included; unrelated model families and training services are | |
| excluded. No new blanket license is asserted over third-party code. | |