SONICS: A Multifaceted Research Term

• SONICS is a polysemous term that primarily denotes a benchmark for synthetic song detection, distinguishing human-composed from AI-generated songs using over 97,000 tracks.
• A key methodological contribution is SpecTTTra, a Spectro-Temporal Tokens Transformer that efficiently models long-range musical structure with up to 38% speedup and reduced memory use compared to conventional models.
• Beyond song detection, SONICS appears in fields like embedded DSP, spatial audio, nuclear spectroscopy, and adversarial machine learning, underscoring its multidisciplinary applications.

SONICS is a polysemous research term. In current music-information-retrieval literature it denotes “Synthetic Or Not – Identifying Counterfeit Songs,” a large-scale benchmark for end-to-end Synthetic Song Detection (SSD) that distinguishes human-composed from AI-generated songs. In other literatures, closely related labels—SONIC, SONIC-O1, SOniCS, SONIC@HORUS, and gyrosonics—identify distinct systems in embedded DSP, multimodal evaluation, biomechanics, nuclear spectroscopy, and auditory stimulation rather than a single technical framework (Rahman et al., 2024, N et al., 16 Jun 2025, Radwan et al., 29 Jan 2026, Mazier et al., 2022, Pickstone et al., 2017, Ghatak et al., 2009).

1. SONICS as a benchmark for end-to-end synthetic song detection

SONICS, expanded as “Synthetic Or Not - Identifying Counterfeit Songs,” was introduced to address a shift from singing voice deepfake detection (SVDD) to end-to-end Synthetic Song Detection (SSD). The motivating distinction is explicit: SVDD asks whether the singing voice is fake while the backing track is typically real, whereas SSD asks whether the song as a whole—vocals, lyrics, instrumentation, arrangement, and style—may be synthetic. The benchmark was designed because prior SVDD datasets relied on real instrumental backings, short clips, limited diversity, and preprocessing pipelines such as voice separation, all of which were considered inadequate for systems such as Suno and Udio that generate complete songs (Rahman et al., 2024).

The dataset scale is central to its role. SONICS contains 97,164 total songs, 4,751 hours of audio, an average duration of 176.03 seconds, 49,074 synthetic songs, and 48,090 real songs. The real songs are curated from YouTube and matched using metadata from the Genius Lyrics Dataset, spanning 9,096 artists. The synthetic side comprises songs from Suno v2, Suno v3, Suno v3.5, Udio 32, and Udio 130, and is subdivided into Full Fake (2,173 songs), Mostly Fake (40,769 songs), and Half Fake (6,132 songs). The paper emphasizes not only scale but also diversity in genres, moods, lyrics metadata, and long-duration structure, while noting that the fake songs are legally shareable because they were created via paid platform subscriptions (Rahman et al., 2024).

2. Long-context modeling and the original SONICS baseline

The principal methodological contribution introduced alongside the benchmark is SpecTTTra, the Spectro-Temporal Tokens Transformer. Its motivation is architectural rather than purely dataset-driven: songs exhibit repeated verses and choruses, evolving arrangement, and cross-section temporal dependencies, so long-range structure should be modeled directly. Conventional CNNs were characterized as local, and ViT-style 2D patch tokenization was characterized as expensive for long audio. SpecTTTra instead tokenizes along time and frequency separately, creating temporal and spectral token streams that attend globally after concatenation (Rahman et al., 2024).

The efficiency argument is formalized through token count. For ViT, the token count is

$$N_{\nu} = \left(\frac{F}{p}\right)\times\left(\frac{T}{p}\right),$$

whereas for SpecTTTra it is

$$N_{\psi} = \left(\frac{F}{f}\right)+\left(\frac{T}{t}\right).$$

This additive rather than multiplicative growth is the main reason the model is presented as time- and memory-efficient for long songs. The three reported variants are SpecTTTra-$\alpha$ with $f=1, t=3$, SpecTTTra-$\beta$ with $f=3, t=5$, and SpecTTTra-$\gamma$ with $f=5, t=7$ (Rahman et al., 2024).

Empirically, SONICS is used with both 5-second and 120-second inputs. On 120-second audio, SpecTTTra-$\alpha$ reports F1 0.97, compared with ConvNeXt at 0.96, ViT at 0.89, and EfficientViT at 0.95. The paper further reports that, for 120-second songs, ViT processes 34 audio/sec at 5.3 GB memory, while SpecTTTra-$\alpha$ processes 47 audio/sec at 3.9 GB memory; this is summarized as 38% faster than ViT with 26% less memory, and compared with ConvNeXt it is reported as a 20% speedup with 67% memory reduction. The broader claim is that detector performance improves substantially when long-form musical structure is retained rather than truncated (Rahman et al., 2024).

3. SONICS in later detector research

SONICS rapidly became a benchmark for both generator-generalization studies and long-audio structural modeling. In MusicDET, it is one of two evaluation datasets and is explicitly used to test cross-generator generalization. MusicDET is trained in a zero-shot setting using real music only, models the likelihood of real music in a frequency-guided normalizing-flow framework, and evaluates detection with Equal Error Rate (EER). The reported SONICS result is 2.89% EER for zero-shot MusicDET, while the class-conditional version reports 0.00% EER. The same table shows Spec-ViT at 1.23% average EER, so the authors do not claim an unconditional numeric win over every baseline; rather, they stress that MusicDET is the only zero-shot method in that comparison and remains competitive without access to generated samples during training (Han et al., 18 May 2026).

The Segment Transformer line treats SONICS as a full-audio benchmark rather than only a short-clip classification task. The earlier Segment Transformer uses beat tracking, downbeat-aware segmentation into 4-bar units, padding or cropping to 48 segments, and a dual-pathway design that combines segment content with a self-similarity representation of track-level structure. On SONICS full-audio detection with MERT features, it reports ACC 0.9992, Prec. 0.9988, Recall 0.9996, F1 0.9992, AUC 0.9999, and Spec. 0.9987 (Kim et al., 10 Sep 2025).

Fusion Segment Transformer (FST) extends that architecture by replacing late concatenation with bi-directional cross-attention and a learnable gate:

$$N_{\psi} = \left(\frac{F}{f}\right)+\left(\frac{T}{t}\right).$$0

with

$$N_{\psi} = \left(\frac{F}{f}\right)+\left(\frac{T}{t}\right).$$1

On SONICS full-audio detection with MERT, FST reports ACC 0.9999, Precision 1.0000, Recall 0.9999, F1 0.9999, AUC 0.9999, and Specificity 1.0000. The same paper reports strong gains for Music2Vec and FXencoder relative to the earlier Segment Transformer, and attributes a substantial part of the gain to musically aligned 4-bar segmentation and explicit content-structure fusion (Kim et al., 20 Jan 2026).

4. Evaluation characteristics, human baseline, and limitations

A recurrent theme in the SONICS literature is that the benchmark is informative but not free of confounds. The original SONICS paper reports a human benchmark on 520 samples with 25-second clips in which human overall F1 is 0.71, while the best AI model in that study, ConvNeXt, reaches 0.92. Humans are reported to struggle especially with Udio 32, where human sensitivity is 0.23. This suggests that the benchmark captures a practically difficult discrimination problem even when automatic detectors perform strongly (Rahman et al., 2024).

At the same time, both dataset and model papers note limitations. The original benchmark reports that YouTube retrieval by title and artist can return incorrect audio, and a manual check of 600 samples suggests about 0.5% noise. It also notes that Udio cannot generate Half Fake songs using real-song lyrics, so HF examples are only from Suno. On the modeling side, the Fusion Segment Transformer paper argues that SONICS may be relatively easy in practice because some synthetic sources, especially Boomy-like generators, leave strong detectable artifacts, and it notes that resampling everything to 16 kHz can create a confound because some real tracks undergo an unnecessary resampling chain for fair comparison with models such as MERT (Kim et al., 20 Jan 2026, Rahman et al., 2024).

These caveats are significant for benchmark interpretation. They do not negate the usefulness of SONICS, but they imply that extremely high accuracy on SONICS should not automatically be read as evidence of complete robustness to future generators, artifact-poor synthesis pipelines, or stylistically broader music corpora. This suggests why later work increasingly emphasizes generator-agnostic or zero-shot formulations and validates beyond SONICS when possible.

5. Related audio and spatial-computing systems using the SONIC/sonics label

Outside synthetic-song detection, the SONIC/sonics naming family appears in several audio, auditory, and spatial-computing systems.

System	Domain	Core contribution
SONIC	Embedded DSP	STM32H753ZI real-time speech enhancement with dual-microphone LMS
Sonic4D	4D scene generation	Training-free spatial audio generation for dynamic 4D scene exploration
SonicMotion	Generative spatial audio	Latent-diffusion FOA generation with dynamic source motion
Gyrosonics	Auditory stimulation	Binaural rotating-sound stimulus linked to ANS modulation
PlugSonic / Sonispace	Interactive authoring	Binaural soundscapes and spatial interfaces for non-expert sound design

SONIC, expanded as “Sound Optimization for Noise In Crowds,” is an embedded real-time speech-enhancement system implemented on the ARM Cortex-M7 STM32H753ZI. It uses a dual-microphone adaptive noise canceller with LMS, a 48 kHz pipeline, DMA-based circular double-buffering, and CMSIS-DSP acceleration. Reported improvements include about 7 dB SNR gain overall, STOI from 0.69 to 0.84, PESQ from 1.69 to 2.55, and MOS 4.47/5 from a live demo with 122 participants (N et al., 16 Jun 2025).

Sonic4D addresses spatial audio generation for immersive exploration of dynamic 4D scenes by combining TrajectoryCrafter for visual generation, MMAudio for monaural audio, GroundingGPT plus monocular depth for 3D source localization, and gpuRIR/Image Source Method simulation for viewpoint-dependent binaural rendering. In a user preference study with 25 participants, Sonic4D reports 89.03% overall preference over MMAudio, with MOS-SLA 4.013 versus 2.322 and MOS-AVSC 3.977 versus 2.418 (Xie et al., 18 Jun 2025).

SonicMotion extends generative spatial audio to first-order Ambisonics with moving sources. It introduces descriptive and parametric variants, a new dataset of simulated spatial audio-caption pairs, and reports that the parametric model improves localization precision from $$N_{\psi} = \left(\frac{F}{f}\right)+\left(\frac{T}{t}\right).$$2, $$N_{\psi} = \left(\frac{F}{f}\right)+\left(\frac{T}{t}\right).$$3, $$N_{\psi} = \left(\frac{F}{f}\right)+\left(\frac{T}{t}\right).$$4 to $$N_{\psi} = \left(\frac{F}{f}\right)+\left(\frac{T}{t}\right).$$5, $$N_{\psi} = \left(\frac{F}{f}\right)+\left(\frac{T}{t}\right).$$6, $$N_{\psi} = \left(\frac{F}{f}\right)+\left(\frac{T}{t}\right).$$7 while keeping semantic/audio quality competitive with AudioLDM 2 and Stable Audio Open (Templin et al., 9 Jul 2025).

Gyrosonics is a different use of the term: a binaural moving-sound stimulus built from a 7/8 Indian Tabla recording, digitally processed to create the perception of sound rotating inside the head. In a pre–post study of 30 volunteers, listening for 9.5 minutes at about 50 dB sensation level was associated with decreased heart rate, reduced systolic blood pressure, and a significantly decreased LF/HF ratio, interpreted by the authors as reduced sympathetic activity and enhanced parasympathetic response (Ghatak et al., 2009).

PlugSonic and Sonispace address authoring rather than detection. PlugSonic is a web- and mobile-based suite for creating, curating, and experiencing binaural soundscapes and sonic narratives using the Web Audio API, the 3D Tune-In Toolkit, and ARKit; its evaluation with inexperienced users reports an average ease score of 6.4/7 and NPS 20 for prototype use (Comunità et al., 2020). Sonispace is a simulated-space interface for sound design that maps emitters, barriers, distance, and material-dependent reflection/transmission into an intuitive spatial workflow, and its preliminary study reports that about 78% of participants enjoyed using the software and about 65% said they would use it if polished (Scarlatos, 2020).

6. Scientific instrumentation and physical measurement

In atmospheric optics, “sonics” refers to sonic anemometers used as the core instrument for long-term turbulence monitoring. A six-year campaign at Dome C, Antarctica, employed up to six modified Applied Technologies Inc. Sx sonic anemometers mounted at heights from 8 m to 45 m, measuring $$N_{\psi} = \left(\frac{F}{f}\right)+\left(\frac{T}{t}\right).$$8 and temperature at 10 Hz to estimate $$N_{\psi} = \left(\frac{F}{f}\right)+\left(\frac{T}{t}\right).$$9 and then $\alpha$0. The study confirms a surface-layer thickness of about 35 m, reports a winter surface-layer seeing contribution of about 1 arcsec, winter coherence times down to about 8 ms, and presents evidence for a second inertial zone at kilometre scales (Aristidi et al., 2018).

In nuclear structure physics, SONIC stands for Silicon Identification Chamber and forms the particle arm of SONIC@HORUS. The setup combines up to 12 silicon $\alpha$1-$\alpha$2 telescopes with a total solid-angle coverage of 9% of $\alpha$3 and a $\alpha$4-ray array of up to 14 HPGe detectors. Its purpose is particle–$\alpha$5 coincidence spectroscopy for low-spin states up to the neutron separation threshold, including reaction-channel identification, excitation-energy reconstruction, event-by-event Doppler correction, and extraction of weak $\alpha$6-decay branching ratios (Pickstone et al., 2017).

In biomechanics, SOniCS is a plugin that couples SOFA with FEniCSx/FFCx so that constitutive models can be modified by one line of code while residual vectors and Jacobian matrices are generated automatically. The paper validates the approach with manufactured solutions, cantilever beams, and FEBio benchmarks, reports machine precision accuracy in several settings, and demonstrates a real-time haptic simulation of a surgical tool in contact with a hyperelastic liver at about 100 FPS on average (Mazier et al., 2022).

7. SONIC in contemporary AI systems beyond music

The SONIC label also appears in machine learning systems unrelated to audio. “Sonic: Fast and Transferable Data Poisoning on Clustering Algorithms” defines SONIC as a genetic poisoning attack that uses FISHDBC as an incremental surrogate to accelerate attacks against HDBSCAN* and related clustering methods. Its main reported advantage is scalability: SONIC is about 27× faster than SlowP on MNIST at 1% poisoning and about 84× faster on CIFAR-10 at 1% poisoning, while preserving strong attack transferability (Villani et al., 2024).

SONIC-O1, expanded as SOcial Natural Interaction Corpus (Omnimodal, v1), is an audio-video benchmark for multimodal LLMs. It contains 4,958 annotations derived from 231 YouTube videos spanning about 60 hours and 13 conversational topics across five broader domains. It evaluates open-ended summarization, multiple-choice question answering, and temporal localization with supporting rationales. The headline result is a 22.6% performance difference in temporal localization between the best closed-source and open-source model families, together with demographic disparities across race, gender, and age groups (Radwan et al., 29 Jan 2026).

A separate 2026 systems paper defines SONIC as “Segmented Optimized Nexus for Information Compression in Key-Value Caching,” a learning-based framework that compresses historical dialogue segments into Nexus tokens for multi-turn LLM deployment. At compression ratios of 80% and 50%, it reports consistent gains over H2O and StreamingLLM on four dialogue benchmarks; on MTBench101, the paper reports an average score improvement of 35.55% over state-of-the-art baselines and a 50.1% speedup over full-context generation (Chen et al., 29 Jan 2026).

Taken together, these usages show that SONICS is not a single discipline-specific term. In current arXiv usage it is most technically established as a benchmark for synthetic-song detection, but the same orthographic family has become a recurring acronymic pattern across audio DSP, spatial sound, atmospheric measurement, biomechanics, multimodal evaluation, adversarial clustering, and LLM systems research.