StillSonicSet: Disambiguating Speech & Sonification

Updated 5 July 2026

StillSonicSet is defined as a synthetic yet acoustically realistic far-field speech dataset for stationary speakers, constructed from LibriSpeech, FSD50K, FMA, and Matterport3D scenes for benchmarking DiT-Flow.
It also represents a real-time, EDM-inspired sonification system that continuously monitors the Swedish supercomputer Kebnekaise by mapping system metrics to layered auditory displays.
The disambiguation emphasizes that despite sharing the same name, the two artifacts serve distinct domains with unique simulation methodologies, evaluation metrics, and practical applications.

StillSonicSet is an ambiguous name in recent arXiv literature. In speech enhancement research, it denotes a synthetic yet acoustically realistic far-field speech dataset for stationary-speaker scenarios, constructed from LibriSpeech, FSD50K, FMA, and 90 Matterport3D scenes, and used as the principal benchmark for DiT-Flow. In a separate sonification paper, the same name denotes a real-time, EDM-inspired sonification system for monitoring the Swedish supercomputer Kebnekaise. The two usages are technically unrelated except for the shared name, so any rigorous discussion requires explicit disambiguation (Cao et al., 23 Mar 2026, Alunno et al., 21 May 2026).

1. Terminological disambiguation

The name has been used for two distinct research artifacts.

Usage of “StillSonicSet”	Domain	Source
Synthetic but acoustically realistic far-field speech dataset for stationary speakers	Speech enhancement, acoustic simulation, robust SE benchmarking	(Cao et al., 23 Mar 2026)
Real-time sonification system for supercomputer monitoring	Sonification, auditory display, HPC monitoring	(Alunno et al., 21 May 2026)

In the speech-enhancement usage, StillSonicSet is introduced as a response to limitations in prior synthetic corpora, especially simplified room impulse response simulation and the relative underrepresentation of stationary speakers in practical settings such as meetings, teleconferencing, and classrooms. In the sonification usage, StillSonicSet is the proposed system for continuous monitoring of an always-running computational environment through indefinitely continuing, stylistically coherent sound (Cao et al., 23 Mar 2026, Alunno et al., 21 May 2026).

2. StillSonicSet as a stationary-source speech dataset

In the DiT-Flow paper, StillSonicSet is defined as a synthetic but acoustically realistic far-field speech dataset designed for speech enhancement under multiple distortions. It is explicitly presented as a variant of SonicSet for stationary speakers, rather than for moving sound sources. The dataset is composed of LibriSpeech for speech, FSD50K for noise and environmental sounds, FMA for music, and 90 Matterport3D scenes for acoustic simulation (Cao et al., 23 Mar 2026).

Its motivating claim is that many earlier synthetic speech-enhancement datasets rely on acoustically simplified simulation, especially shoebox-style or image-source-method assumptions, and therefore do not adequately represent complex room geometries, varied surface materials, or natural occlusions such as furniture and architectural structure. StillSonicSet is intended to narrow that realism gap while retaining the scale advantages of synthetic data. The target use case is not generic source motion, but the stationary-source conditions common in telecommunication and meeting scenarios (Cao et al., 23 Mar 2026).

The distortions directly modeled within StillSonicSet are reverberation, background noise/music, and Opus codec compression. The paper treats codec artifacts as a distinct source of degradation because practical communication pipelines often involve low-delay compression before enhancement. This makes StillSonicSet not merely a reverberant-speech corpus, but a benchmark for compound distortions (Cao et al., 23 Mar 2026).

3. Construction, simulation lineage, and acoustic ingredients

StillSonicSet is derived from the earlier SonicSet/SonicSim framework, which was introduced for moving sound source scenarios using LibriSpeech, FSD50K, FMA, and 90 scenes from Matterport3D. SonicSim itself is a customizable simulation platform based on Habitat-sim, intended to model indoor acoustics with occlusion, complex room geometry, heterogeneous material properties, and source motion (Li et al., 2024).

Within that lineage, StillSonicSet reuses the RIRs provided in SonicSet and converts moving-source simulation into a stationary-source benchmark by discretizing the moving RIR trajectories and taking responses at fixed positions. For each scene, the source and microphone placements are randomly generated, and the initial position of each speech source and the position of noise sources are placed within a 1–8 meter radius of the microphone. To simulate stationary speakers, each speech utterance is convolved with the RIR from one fixed position only, and the volume is not normalized across positions, so level changes follow source–microphone distance naturally (Cao et al., 23 Mar 2026).

The speech component is assembled by randomly selecting three speech utterances from different speakers from LibriSpeech, downsampling them to 8 kHz, and convolving them with three distinct RIRs. Background noise and music are randomly selected from FSD50K and FMA and convolved with other RIRs from the same acoustic environment, so speech and non-speech share consistent spatial characteristics within a scene. The codec component is added using Opus through the opuslib Python package, with 8 kHz audio, 16-bit precision, 20 ms encoding frames, encoder complexity 10, and target bitrate randomly sampled from 30 kbps to 40 kbps (Cao et al., 23 Mar 2026).

The realism claim is therefore architectural rather than documentary: StillSonicSet is not a real-recorded corpus, but a synthetic corpus whose room simulation incorporates complex geometry, surface-material heterogeneity, occlusions, distance-dependent level variation, and scene-consistent spatialization (Cao et al., 23 Mar 2026).

4. Benchmark protocol and dataset scale

StillSonicSet serves as the primary dataset for the core DiT-Flow experiments. The paper reports experiments on two versions of the dataset: one containing only reverberant speech, and another augmented with noise and compression distortions. The benchmark evaluations are reported under four conditions: Reverb only, Noise only, Reverb + Noise, and Reverb + Noise + Codec-Compression. The last condition is intended to emulate realistic teleconferencing pipelines by applying codec compression after reverberation and additive interference (Cao et al., 23 Mar 2026).

For the main experiments, the reported splits are:

training: 50,000 utterances, approximately 90 hours
validation: 8 hours
test: 8 hours

The evaluation metrics reported on StillSonicSet are PESQ, ESTOI, LSD, DNSMOS P.835 (SIG, BAK, OVRL), and speaker similarity based on WavLM-based cosine similarity. The benchmark is therefore designed to assess intelligibility, spectral fidelity, perceptual quality, background suppression, and speaker preservation within the same experimental setting (Cao et al., 23 Mar 2026).

This benchmark design distinguishes StillSonicSet from a purely reverberant or purely noisy corpus. The dataset is used to test robustness not only to single distortions but also to the interaction between reverberation, interference, and codec artifacts.

5. Reported empirical role in DiT-Flow

StillSonicSet is used to compare SGMSE, StoRM, and DiT-Flow under matched training and evaluation conditions. The most emphasized condition is Reverb + Noise + Codec-Compression, which the paper presents as the hardest and most realistic teleconferencing-like setup. In that condition, the noisy input has PESQ 1.126, ESTOI 0.312, LSD 8.293, SIG 1.545, BAK 1.494, OVRL 1.277, and speaker similarity 0.779. DiT-Flow improves these to 1.389, 0.458, 4.506, 3.301, 3.723, 2.906, and 0.880, respectively, while StoRM remains higher on BAK at 3.969 (Cao et al., 23 Mar 2026).

The paper’s broader conclusion is that DiT-Flow is the most robust model under multi-distortion conditions, especially in the compound setting combining reverberation, noise, and codec compression. It does not dominate every metric in every condition, but it is reported as providing the strongest overall balance across signal naturalness, intelligibility, spectral fidelity, and speaker preservation (Cao et al., 23 Mar 2026).

StillSonicSet’s significance is reinforced by transfer experiments to real-recorded datasets. For LibriCSS, DiT-Flow trained on WSJ0+Reverb yields SIG 2.938, BAK 3.902, OVRL 2.476, and Spk Sim 0.917, whereas training on StillSonicSet yields SIG 2.935, BAK 3.950, OVRL 2.503, and Spk Sim 0.928. For RealMAN, the corresponding comparison is 2.754 / 3.413 / 2.184 / 0.885 versus 2.911 / 3.684 / 2.402 / 0.896. The larger gains on RealMAN are interpreted in the paper as evidence that StillSonicSet better matches complex real acoustics than the simpler synthetic reverberation baseline (Cao et al., 23 Mar 2026).

6. Scope boundaries, common misreadings, and the separate sonification usage

Several boundaries are important. First, StillSonicSet is synthetic, not real-recorded. Its claim is “synthetic yet acoustically realistic,” not documentary realism. Second, it should not be conflated with SonicSet itself: SonicSet is a moving-source benchmark, whereas StillSonicSet is a stationary-source adaptation derived from the same simulation lineage. Third, the five unseen distortions discussed elsewhere in the DiT-Flow paper—clipping, bandwidth limitation, codec loss (MP3/OGG), packet loss, and wind noise—belong to the URGENT adaptation study, not to StillSonicSet itself (Cao et al., 23 Mar 2026).

A further source of confusion is the unrelated sonification paper that uses the same name. In that work, StillSonicSet is not a speech dataset at all, but a system for real-time sonification of the activity of the Swedish supercomputer Kebnekaise. It monitors data exposed by Slurm, specifically \procs, \memusage, and \IB-tx, receives a new batch every 15 seconds, maps the machine’s 10 partitions to 10 layers of an EDM texture, and uses a round-robin foregrounding strategy to manage masking and cognitive overload. That StillSonicSet is a monitoring sonification environment rather than a speech-enhancement benchmark (Alunno et al., 21 May 2026).

The shared name therefore denotes two different objects in two different subfields. In the speech literature, StillSonicSet refers to a stationary-source, acoustically realistic synthetic benchmark for far-field speech enhancement under reverberation, background noise/music, and Opus codec compression. In sonification, it refers to a continuous auditory display system for supercomputer activity. The two usages should be cited and interpreted separately (Cao et al., 23 Mar 2026, Alunno et al., 21 May 2026).