SPRSound 2024: Advances in SELDDE

• SPRSound 2024 is a sound event localization, detection, and distance estimation system that leverages a ResNet-Conformer backbone with SALSA features and a robust augmentation pipeline.
• It integrates advanced squeeze-and-excitation attention and a two-step distance scaling procedure to enhance localization accuracy and manage regression errors from FOA recordings.
• Evaluation on the STARSS23 dataset shows significant improvements in F1-score and reduced SELDDE errors, underscoring its practical potential in real-world acoustic sensing.

Searching arXiv for “SPRSound 2024” and closely related terms to ground the article in papers. SPRSound 2024 denotes a 2024 sound-event localization, detection, and distance-estimation configuration developed for Task 3 of the DCASE 2024 Challenge, specifically the audio-only Sound Event Localization and Detection with Source Distance Estimation (SELDDE) setting. Its defining elements are a ResNet-Conformer backbone augmented with squeeze-and-excitation attention, Spatial Cue-Augmented Log-Spectrogram (SALSA) features, a two-step distance-scaling procedure for the regression target, and a training pipeline that combines real STARSS23 data with SpatialScaper-generated synthetic First-Order Ambisonics (FOA) recordings and spectrogram-level augmentation (Yeow et al., 2024).

1. Problem setting and representational target

The system addresses only the audio-only SELD with SDE task. In this setting, the objective is joint sound event localization, detection, and distance estimation from FOA recordings, with performance reported on the evaluation subset of the STARSS23 dataset through macro-averaged, class-dependent metrics and an aggregated SELDDE error (Yeow et al., 2024).

The output representation is the multi-ACCDDOA vector of dimension $(3 \times C)$ per time step, where $C$ is the number of event classes. The output head is a $1 \times 1$ convolution followed by $\tanh(\cdot)$, and training uses Mean Squared Error between predicted and ground-truth multi-ACCDDOA targets, with auxiliary duplicating Permutation-Invariant Training as in Shimada et al. (Yeow et al., 2024).

A central representational choice is the replacement of commonly used log-mel spectra with SALSA. For each of the four FOA channels $(W, X, Y, Z)$, the system computes a log-magnitude spectrogram,

$$S_c(t,f)=20\,\log_{10}\bigl(|\mathrm{STFT}\{x_c\}(t,f)|+\epsilon\bigr),$$

and supplements it with three inter-channel Level Differences relative to the omnidirectional channel $W$,

$$\mathrm{ILD}_i(t,f)=20\log_{10}\!\bigl(\tfrac{|X_i(t,f)|}{|W(t,f)|}\bigr), \quad i\in\{X,Y,Z\}.$$

These are concatenated into a 7-channel tensor comprising four spectral channels and three spatial ILD channels. Above 9 kHz the linear frequency bins are compressed linearly so that the final feature map has 200 frequency bins (Yeow et al., 2024).

2. Architectural organization

The input is segmented into non-overlapping 5 s windows, sampled at 24 kHz and converted to SALSA features of shape $$(7 \text{ channels} \times 400 \text{ time-bins} \times 200 \text{ freq-bins})$$. The encoder is ResNet-18-style, with four residual stages; each stage contains two $3 \times 3$ convolution layers, with channels doubling per stage from $C$0, using ReLU and batch normalization (Yeow et al., 2024).

A distinctive structural decision is that pooling is applied only along the frequency axis after each residual block through AvgMaxPool over all frequency bins, thereby preserving full temporal resolution. The encoder is followed by three stacked Conformer modules, each composed of a Feed-Forward layer, a Multi-Head Self-Attention layer, a convolutional module with depthwise and pointwise convolution, and a second Feed-Forward layer, with residual connections wrapping each sub-module. No pooling is used inside the conformers. After the last conformer, AvgMaxPool is applied along the time and frequency dimensions before the output head (Yeow et al., 2024).

The squeeze-and-excitation mechanism is integrated within the ResNet encoder in three variants: cSE, sSE, and SCSE. For an input feature map $C$1, channel-SE performs spatial squeeze via global average pooling,

$C$2

followed by a two-layer gating operation with reduction ratio $C$3,

$C$4

and channel-wise rescaling. Spatial-SE uses a $C$5 convolution to produce a single-channel mask

$C$6

which is then used for spatial rescaling. Concurrent SCSE combines channel and spatial excitation by element-wise summation of the respective outputs. In the reported configuration, sSE blocks are inserted in the stem of each ResNet stage before any pooling, SCSE blocks are inserted after each residual stage, and each cSE uses $C$7 (Yeow et al., 2024).

3. Feature engineering, augmentation, and distance scaling

The feature pipeline is coupled to an augmentation strategy intended to increase both spatial diversity and class balance. Audio Channel Swapping randomly permutes the four FOA channels within a training example, thereby creating new DOA representations while preserving room reverberation and diffuse characteristics; it is applied only to real STARSS23 data, at fixed probability throughout training. SpatialScaper synthetic data contribute approximately 18 h of FOA recordings synthesized using real SRIRs whose source-receiver distance distributions match STARSS23, spanning 0.04 m to 7.64 m. The synthesis process up-samples under-represented classes such as “knock” and “doorCupboard,” limits polyphony to $C$8, and uses 8 distinct rooms (Yeow et al., 2024).

At the spectrogram level, the training pipeline applies SpecAugment with time-mask and frequency-mask operations, Mixup along time and frequency, and random frequency shifting. The magnitudes of these operations are annealed via a cosine schedule as the learning rate decays. This design makes the system not merely an architectural modification but a coordinated representation-and-augmentation stack (Yeow et al., 2024).

Distance estimation is explicitly normalized so that large-distance errors do not dominate the loss. Ground-truth distances $C$9 are first standardized,

$1 \times 1$0

using the empirical mean and standard deviation of all distances in training, and are then scaled into $1 \times 1$1,

$1 \times 1$2

At inference, these two steps are inverted to recover distances in meters. The stated rationale is that, within the MSE objective, all distance errors are thereby bounded equally (Yeow et al., 2024).

4. Optimization protocol and submitted variants

Optimization uses Adam with a peak learning rate of $1 \times 1$3 and a “Transformer” learning-rate schedule with warmup and inverse-square-root decay as in Vaswani et al. The batch size is 32. Training proceeds in two phases: 200 epochs on STARSS23 plus SpatialScaper synthetic data with all augmentations, followed by 50 epochs of fine-tuning on STARSS23 real data only, with augmentations still applied. Model checkpointing selects the model with the highest validation location-dependent F1-score (Yeow et al., 2024).

The reported system family comprises a baseline, a retrained enhanced baseline, and four submitted variants. The baseline is FOA, ResNet-Conformer, no SE blocks, with mel-spectrogram input. The “BL$1 \times 1$4” system retrains the baseline using SALSA, distance scaling, and all augmentations. Systems A–D all use SALSA and the broader training pipeline, but differ in the insertion of SE blocks. System A uses all SE blocks; System B omits the final SCSE; System C removes SCSE from the ResNet stages; System D uses no SE blocks at all (Yeow et al., 2024).

This organization is significant because it separates gains due to representation and data generation from gains due to SE-based attention. A plausible implication is that SPRSound 2024 should be understood as a composite methodology rather than as a single architectural trick.

5. Evaluation, quantitative performance, and interpretation

Evaluation on the STARSS23 development test split uses four reported metrics: the location-dependent F1-score $1 \times 1$5 with $1 \times 1$6 tolerance, class-dependent localization error $1 \times 1$7, class-dependent relative distance error $1 \times 1$8, and the aggregated SELDDE error

$1 \times 1$9

The metrics are macro-averaged over classes with jackknife estimation (Yeow et al., 2024).

System	Configuration	Reported metrics
Baseline	FOA, ResNet-Conformer, no SE, mel-spectrogram	$\tanh(\cdot)$0, $\tanh(\cdot)$1, $\tanh(\cdot)$2, $\tanh(\cdot)$3
BL$\tanh(\cdot)$4	SALSA + distance scaling + all augmentations	$\tanh(\cdot)$5, $\tanh(\cdot)$6, $\tanh(\cdot)$7, $\tanh(\cdot)$8
System A	all SE blocks	$\tanh(\cdot)$9, $(W, X, Y, Z)$0, $(W, X, Y, Z)$1, $(W, X, Y, Z)$2
System B	omit the final SCSE	$(W, X, Y, Z)$3, $(W, X, Y, Z)$4, $(W, X, Y, Z)$5, $(W, X, Y, Z)$6
System C	no SCSE in ResNet stages	$(W, X, Y, Z)$7, $(W, X, Y, Z)$8, $(W, X, Y, Z)$9, $$S_c(t,f)=20\,\log_{10}\bigl(|\mathrm{STFT}\{x_c\}(t,f)|+\epsilon\bigr),$$0
System D	no SE blocks at all	$$S_c(t,f)=20\,\log_{10}\bigl(|\mathrm{STFT}\{x_c\}(t,f)|+\epsilon\bigr),$$1, $$S_c(t,f)=20\,\log_{10}\bigl(|\mathrm{STFT}\{x_c\}(t,f)|+\epsilon\bigr),$$2, $$S_c(t,f)=20\,\log_{10}\bigl(|\mathrm{STFT}\{x_c\}(t,f)|+\epsilon\bigr),$$3, $$S_c(t,f)=20\,\log_{10}\bigl(|\mathrm{STFT}\{x_c\}(t,f)|+\epsilon\bigr),$$4

The numerical pattern shows that the major improvement over the baseline is not attributable solely to SE blocks. BL$$S_c(t,f)=20\,\log_{10}\bigl(|\mathrm{STFT}\{x_c\}(t,f)|+\epsilon\bigr),$$5 already improves the baseline from $$S_c(t,f)=20\,\log_{10}\bigl(|\mathrm{STFT}\{x_c\}(t,f)|+\epsilon\bigr),$$6 to $$S_c(t,f)=20\,\log_{10}\bigl(|\mathrm{STFT}\{x_c\}(t,f)|+\epsilon\bigr),$$7 and reduces $$S_c(t,f)=20\,\log_{10}\bigl(|\mathrm{STFT}\{x_c\}(t,f)|+\epsilon\bigr),$$8 from $$S_c(t,f)=20\,\log_{10}\bigl(|\mathrm{STFT}\{x_c\}(t,f)|+\epsilon\bigr),$$9 to $W$0, while System D, despite using no SE blocks at all, reaches $W$1 and $W$2. At the same time, the best reported result is System A with $W$3, $W$4, $W$5, and $W$6. The report summarizes the outcome as halving the aggregated SELDDE error and improving localization F1 from 13.1% to 33.9% (Yeow et al., 2024).

A common misconception would be to treat distance estimation as the dominant beneficiary of the method. The reported numbers indicate that the strongest changes are in localization-sensitive metrics, particularly $W$7 and $W$8, whereas $W$9 changes more modestly across the submitted systems. This suggests that the representation and attention design chiefly strengthen joint localization and class assignment, with distance estimation improved but not transformed to the same degree.

6. Limitations, open questions, and relation to broader acoustic sensing

The stated limitations are threefold. First, the ablation of SE block placement remains inconclusive, and optimal insertion points are not fully explored. Second, distance outputs are constrained by the training distribution of 0.04 m to 7.64 m. Third, evaluation is limited to FOA format and the STARSS23 subset, so real-world mismatch could still exist. The listed future directions are a systematic study of SE block placement and alternative attention mechanisms, development of novel handcrafted or learned spatial-distance features beyond SALSA, extension to multi-modal audio-plus-visual SELDDE, and ensembling with more diverse synthetic data including moving sources and microphone arrays (Yeow et al., 2024).

In a broader acoustics context, SPRSound 2024 belongs to a wider family of acoustic inference systems in which localization depends on the joint treatment of propagation, sensor geometry, and inverse modeling. The South Pole Acoustic Test Setup provides an example from a different application domain: SPATS used four strings deployed in IceCube boreholes, monitored the 10 to 100 kHz range, reconstructed source vertices from sensor positions and hit times, and identified reconstructed transients with re-freezing IceCube holes and Rodriguez wells rather than unexplained sources (Karg, 2010). In a companion study, SPATS formalized reconstruction through global positioning equations, used all $$\mathrm{ILD}_i(t,f)=20\log_{10}\!\bigl(\tfrac{|X_i(t,f)|}{|W(t,f)|}\bigr), \quad i\in\{X,Y,Z\}.$$0 four-sensor subsets to obtain ensembles of reconstructed vertices, and derived a flux limit for high-energy neutrinos after observing zero neutrino-like events in a vetoed background-free region over 245 days (Berdermann, 2010).

The two programs are methodologically distinct: SPRSound 2024 operates on FOA scene analysis with neural sequence modeling, whereas SPATS studies acoustic propagation and transient reconstruction in Antarctic ice. Even so, the comparison is technically instructive. It suggests that advanced learned representations such as SALSA and multi-ACCDDOA occupy, for room-scale SELDDE, a role analogous to calibrated propagation models and time-of-arrival reconstruction in sparse scientific arrays: both are mechanisms for converting structured acoustic measurements into estimates of source class, position, and event validity.