Class-Aware SDR Improvement (CA-SDRi)

• Class-Aware SDR Improvement (CA-SDRi) is a metric that evaluates spatial semantic segmentation by averaging per-class SDR enhancements and penalizing misclassifications.
• The methodology employs enriched audio features, agent-based label correction, and dataset refinement to improve both source separation and classification accuracy.
• Quantitative analysis shows CA-SDRi can boost performance by up to 14.7%, demonstrating significant gains in handling complex sound scenes.

Class-Aware SDR Improvement (CA-SDRi) is a metric and system design paradigm used to evaluate and drive performance advances in spatial semantic segmentation of sound scenes (S5), where the joint objectives of source separation and sound event classification must be satisfied. CA-SDRi provides a unified, class-balanced measure of how much a system improves the signal-to-distortion ratio (SDR) for each class, penalizing both separation and classification errors. State-of-the-art systems leverage architectural innovations, enriched audio features, agent-based error correction, and dataset optimization to maximize CA-SDRi, especially in the context of the DCASE 2025 Task 4 evaluation protocol.

1. Definition and Rationale of CA-SDRi

CA-SDRi extends conventional SDR improvement metrics by explicitly coupling source separation quality with classification correctness on a per-class basis. Formally, for $C$ target sound event classes, CA-SDRi is computed as

$\mathrm{CA\textsubscript{SDRi}} = \frac{1}{C} \sum_{c=1}^{C} \mathrm{SDRi}_c,$

where $\mathrm{SDRi}_c$ is the mean SDR improvement for all sources of class $c$ (Park et al., 26 Jun 2025). This stands in contrast to a global SDRi that simply averages over all sources regardless of class, a practice that risks underrepresenting classes with few examples. In recent extensions, the metric has been further refined (denoted as CASA-SDR or CA-SDRi with penalties) to account for mistaken classifications by penalizing mis-classified or cross-contaminated output sources (Mishra et al., 10 Nov 2025). The link between class awareness and SDR improvement allows CA-SDRi to serve as both a diagnostic and optimization target for real-world polyphonic source separation systems.

2. Mathematical Formulation and Metric Comparisons

CA-SDRi can be viewed as an aggregation of per-class SDR improvements, ensuring class balance and robust performance assessment:

• For a mixture with ground-truth class set $\mathcal{C}$ and predicted class set $\hat{\mathcal{C}}$, the dominant CA-SDRi definition in DCASE 2025 Task 4 is

$$\mathrm{CA\text{-}SDRi} = \frac{1}{|\mathcal{C} \cup \hat{\mathcal{C}}|} \sum_{k \in \mathcal{C} \cup \hat{\mathcal{C}}} P_k,$$

with

$$P_k = \begin{cases} \mathrm{SDRi}_k, & \text{if } k \in \mathcal{C} \cap \hat{\mathcal{C}}, \ 0, & \text{otherwise.} \end{cases}$$

• Only true-positive classes—those detected and correctly labeled—contribute SDR improvement; false negatives and false positives assign zero (Kwon et al., 17 Sep 2025).
• Recent proposals (CASA-SDR) find the best permutation of system outputs to maximize total SDR (matching in SDR-space, not by class), then explicitly penalize misclassifications, using either input-level or output-level penalties, with options for non-TP or error-based penalty applications (Mishra et al., 10 Nov 2025).

The adoption of CA-SDRi avoids interpretation errors where a system's poor classification could be misattributed to poor separation, ensuring that end-to-end assessment faithfully represents both dimensions.

3. System Design Strategies for Maximizing CA-SDRi

Recent high-performing systems (notably for DCASE 2025 Task 4) employ a coordinated strategy:

1. Feature Enrichment: Integration of advanced audio features such as spectral roll-off and chroma, in addition to conventional log-mel spectrograms, to capture class-distinctive cues overlooked by traditional representations. For an input $x$, features are extracted as follows:

• Spectral roll-off per frame $t$: $r_t$ is the frequency bin index such that the cumulative energy exceeds 85%,
• Chroma features: energy is summed into 12 pitch bins for each time frame.

These features are projected into 256-dimensional embeddings (via MLPs or CNNs) and then concatenated to a 768-dimensional vector, which feeds the classification head (Park et al., 26 Jun 2025).

2. Agent-Based Label Correction: Implementation of an error-correcting agent to suppress false positives:

• Candidates are the top-3 audio tagging scores or any exceeding threshold $\tau=0.5$.
• For each, perform class-conditioned separation, re-classify the result, and retain the class only if it is still top-ranked or exceeds threshold on its own extracted output.
• The corrected set is then used for final separation, and CA-SDRi is measured accordingly.

This process empirically reduces false positives and systematically improves CA-SDRi, delivering measurable percentage gains over baseline (Park et al., 26 Jun 2025).

3. Dataset Refinement and Class Balancing:

• Removal of source clips shorter than 1.5 s and manual curation to eliminate perceptually out-of-class examples.
• Augmentation of underpopulated classes using external corpora (e.g., AudioSet), thus ensuring balanced class coverage, which is crucial for class-averaged metrics.

4. Quantitative Impact and Ablation Analyses

Empirical studies on Task 4 of DCASE 2025 show the additive benefit of each design component in boosting CA-SDRi:

System Component	CA-SDRi (dB)	Relative Gain
Baseline (official checkpoint)	11.088	—
+ Agent-based correction	11.244	+1.4%
+ Dataset refinement	12.306	+11%
+ Spectral roll-off feature	12.328	+0.18%
+ Chroma feature	12.426	+0.98%
+ Roll-off + Chroma	12.532	+1.81%
Final 4-model ensemble	12.721	+14.7%

The cumulative effect is a 14.7% relative increase in CA-SDRi from the official baseline to the ensemble system, with absolute performance rising from 11.088 dB to 12.721 dB (Park et al., 26 Jun 2025). Additional studies corroborate similar CA-SDRi boosts with multi-stage, self-guided frameworks integrating universal sound separation, iterative re-classification, and clue-conditioned extraction (Kwon et al., 17 Sep 2025).

5. Extensions: CA-SDRi with Penalties and Performance Analysis

Recent metric analyses identify limitations of the original CA-SDRi metric, especially in disambiguating the sources of performance drop between separation and classification:

• Permutation-insensitive SDR matching, followed by explicit classification error accounting, forms the basis of improved metrics (CASA-SDR or penalized CA-SDRi).
• Penalties are introduced for misclassifications, either at the input level (reflecting the prominence of the source in the mixture) or output level (reflecting the system's separation fidelity for the wrongly labeled source). Non-TP and EB (error-based) variants control for penalty multiplicity and strictness (Mishra et al., 10 Nov 2025).
• Comparative studies show that CA-SDRi (with penalties) unambiguously separates separation failures from labeling errors, avoids artifacts caused by class-assignment permutations, and allows tuning penalty schemes to target specific application requirements.

6. Practical Implications for System Design and Evaluation

CA-SDRi serves as the primary target for optimization in modern spatial semantic segmentation systems due to its sensitivity to both separation and recognition accuracy. Notable implications include:

• Feature and architecture choices should maximize class-distinct structure in embeddings.
• Aggressive post-processing (e.g., agent/consistency-based correction) is essential to suppress systematic labeling errors, which otherwise zero out or penalize CA-SDRi irrespective of separation quality.
• Dataset curation is critical: class imbalance or label noise directly impairs the metric by pushing down the per-class average.
• Improvements in CA-SDRi translate to real-world gains in intelligibility and event recognition in autonomous auditory systems, with reported improvements of 4–5 dB being significant for practical deployment scenarios (Kwon et al., 17 Sep 2025).

7. Limitations and Considerations in Metric Usage

CA-SDRi, while robust, exposes several practical and conceptual limitations:

• It can mask the details of individual error type distribution (separation vs. classification) unless penalty-based refinements are employed.
• In high-SNR or high-separation settings, mis-classification penalties may dominate, potentially driving scores negative and necessitating careful tuning for fair comparisons (Mishra et al., 10 Nov 2025).
• A plausible implication is that metric selection and implementation details (e.g., the precise penalty scheme or class-balance normalization) can have substantial effect on system rankings.
• Adoption of CA-SDRi, especially with penalty variants, provides a transparent and tunable joint metric, increasingly favored in DCASE and related S5 challenges.

CA-SDRi thus encapsulates the state-of-the-art approach to evaluating and improving spatial sound scene understanding, facilitating advances in both algorithm design and benchmark methodology (Park et al., 26 Jun 2025, Mishra et al., 10 Nov 2025, Kwon et al., 17 Sep 2025).