DaCAS-3 on deliverables
This is a permanent document (pdoc) v0.21 serving as a tracking mechanism for defining Data-driven Channel-Adaptive Spatial Audio (DaCAS) example solutions and their deliverables. The document provides the structural framework that will feed into TS 26.533.
The document establishes a standardized template for documenting DaCAS example solutions, ensuring consistency across contributions from different sources. It functions as a living document where agreed contributions will be integrated over time.
The document provides a comprehensive skeleton structure for future example solution contributions. This template ensures all solutions are documented with consistent information categories.
2.1.3.1 Input Interfaces
- Placeholder for describing input interface specifications
- Currently TBD
2.1.3.2 Data
- Distinguishes between two solution types:
- Neural-network-based solutions: Requires description of training data used to develop the solution
- Signal-processing-based solutions: Requires description of necessary data (e.g., geometric data of target devices)
- Currently TBD
2.1.3.3 Processing Procedure
- Should include descriptions of:
- Offline signal processing procedures
- Online signal processing procedures
- Model structure (for ML-based solutions)
- Currently TBD
The document establishes different requirements based on solution type:
For signal processing-based solutions or modules:
- Source code of the algorithm must be included in the example solution package
For neural-network-based solutions or modules:
- Final checkpoint (trained weights) must be provided
- Inference code must be provided
- Training code and intermediate checkpoints are optional
2.1.5.1 Usage Guideline
- Should contain instructions on how to compile, run, and test the code
- Currently TBD
2.1.7.1 Objective Evaluation Results
- Placeholder for objective metrics and results
- Currently TBD
2.1.7.2 Subjective Evaluation Results
- Optional section for subjective test results
- Currently TBD
The main contribution is establishing a unified documentation framework that:
1. Accommodates both neural-network-based and signal-processing-based solutions
2. Ensures reproducibility through mandatory source code/model sharing
3. Requires comprehensive evaluation (objective and optionally subjective)
4. Addresses practical implementation aspects (usage guidelines, test vectors)
5. Includes optional performance analysis sections (complexity, delay)
6. Incorporates legal framework considerations
The template explicitly supports two paradigms:
- Data-driven (ML-based) approaches: With specific requirements for trained models and inference code
- Traditional signal processing approaches: With requirements for algorithmic source code and geometric data
Clear distinction between mandatory and optional deliverables ensures baseline comparability while allowing flexibility for additional information.
The document builds upon:
- S4-250963: "Proposal for DaCAS deliverables" (Nokia)
- S4aA260009: "[DaCAS] Example solution deliverables" (Bytedance, Beijing Xiaomi Mobile Software)
This is a template document (v0.21) with all technical sections marked as TBD, awaiting contributions from participants to populate the framework with actual example solutions.