Report from Audio SWG teleconference on DaCAS (9 December 2025)
The Audio SWG held a teleconference on DaCAS (Diverse audio CApturing system for Smartphone devices) with 19 participants for 1 hour. Three input documents were discussed, all focused on defining the evaluation approach, specification methodology, and deliverables for DaCAS example solutions.
Key Proposals:
- Self-evaluation approach for example solutions rather than cross-evaluation
- High-level documentation guidelines for proponents
- Informal assessment against requirements in TS 26.261
- Updated work plan with two timeline options depending on evaluation procedure chosen
Timeline Proposals:
- Table 1: Faster timeline with self-evaluation only
- Table 2: Extended timeline including cross-evaluation
- Key milestones: submission of example solutions in Montreal, specification work following
Discussion Points:
- Mixed support: some agreement on self-evaluation concept, but reservations about completely skipping cross-evaluation
- Questions on deliverable package definition (executable vs source code)
- Concerns about timeline dependencies on deliverable agreements
- Clarification needed on what "informal evaluation" means in practice
- Need to distinguish between demonstrating feasibility vs determining suitability of solutions
Outcome: Document noted; offline discussions encouraged to provide potential revision
Key Proposals:
- Two-tier approach for example solutions:
- All evaluations and documentation included in informative Annex of TS 26.533
- Example solutions additionally meeting requirements from TS 26.261 specified in TS 26.533 (status TBD)
Documentation Requirements:
- Algorithmic description with sufficient detail for implementation
- Optional elements:
- Adaptation considerations for commercial devices
- Support for IVAS PI (Parametric Information) data
- Integration with commercial processing pipelines
- Device limitations and generalization capabilities
Discussion Points:
- Questions on required detail level: pseudocode vs source code
- Clarification that last three bullets are optional/invitational
- Relationship to ATIAS Phase 3 work
- Distinction between two categories:
1. Feasible example solutions (demonstration via self-evaluation)
2. Endorsed solutions meeting TS 26.261 requirements (requiring testing)
- Suggestion to share draft updates via SA4 reflector for broader input
Outcome: Document noted; revision expected with broader consultation
Key Proposals for Deliverable Package:
For Neural Network-based Solutions:
- Code package including:
- Training code
- Inference code
- Model checkpoints
- Rationale: enable retraining for different devices beyond target device
For DSP-based Solutions:
- Algorithmic description allowing reimplementation
General Requirements:
- Legal framework submission (depending on solution specifics)
- Sufficient details to convert raw microphone recordings to IVAS formats
Discussion Points:
- Strong concerns about mandating training code and checkpoints:
- Risk of providing false impression of quality transferability
- Training typically tailored to specific devices
- Cannot guarantee similar quality on different devices
- May be unnecessary if solution generalizes without retraining
- Distinction between providing model (equivalent to DSP description) vs full training pipeline
- Legal framework requirements too general and need clarification
- Fundamental difference in reproducibility between DSP descriptions (anyone can reimplement) and NN solutions
Outcome: Document noted; significant concerns raised about mandatory training code requirements