Audio Representation Learning

When does distribution alignment help a frozen foundation-model embedding generalize across acoustic domains? For cross-domain mosquito-species classification we report a strength-monotonic law: the stronger an encoder is on the target task, the more its unseen-domain generalization relies on a distribution-alignment (MMD) term, and the more it is harmed by domain-rebalanced sampling. Across four encoder families and a within-encoder HuBERT layer sweep (n=8), the rebalancing leg orders exactly with encoder strength (Spearman -1.000), while the MMD-benefit leg is monotonic within each stream and -0.857 pooled; fixing architecture and varying only representation strength flips the rebalancing effect from benefit to collapse. The law is actionable: a single MMD term is the sole lever on a strong encoder, so we reduce the field's default recipe to a frozen Perch 2.0 embedding, a lightweight probe, cross-entropy, one MMD, and input augmentation. The reduced recipe stays within seed noise of the full composite (BA_unseen 0.299+/-0.006 vs. 0.307+/-0.014). As boundary conditions of the same law, three community defaults (backbone fine-tuning, multi-modal fusion, and domain rebalancing) each hurt unseen-domain accuracy under a leave-domain protocol, shown with single-variable, multi-seed evidence. We present a mechanism and the recipe it explains, not a leaderboard entry.

Neural audio synthesis models like the Realtime Audio Variational autoEncoder (RAVE) achieve impressive genera tion quality, yet how their internal representations encode musical features remains poorly understood. We present a systematic layer-wise and cross-layer cluster analysis of RAVE decoder activations across three models trained on different musical domains, tested with four stimulus types. We then evaluate architectural generalization with a general purpose EnCodec model. For RAVE, we find that synthetic stimuli are encoded well across models and audio features (pitch |\r{ho}|=0.45, 5.1x the null, BPM |\r{ho}| = 0.76, 8.6x the null). These results are reduced but still substantively apparent when using natural audio (mean across features |\r{ho}|=0.25, 2.8x the null). Natural audio sees a stronger encoding when nonlinear probes are used (mean across features R2=0.56, 18x the null, +0.152 nonlinear gain over the linear probe R2). Encoding strength varies throughout the layers of the decoder and an increased ability to joint-encode in the middle layers is seen across all audio features (\b{eta}2 all negative, p < 0.05). The general purpose EnCodec decoder also sees similar strong synthetic responses across audio features, similar nonlinear gains for natural audio joint encoding and similar depth profiles. We find the best cross-layer cluster improves the strength (r = 0.65, p = 0.006) and prevalence (r = 0.75, p = 0.001) of BPM encoding when compared against the best whole layers within the same section, with no effect for joint encoding. These findings advance the interpretability of neural audio models and inform targeted control strategies for neural synthesis.

Speech-pretrained models have shown promise in animal bioacoustics, but the factors governing their cross-species transfer remain poorly understood. We evaluate 15 frozen encoders spanning monolingual and multilingual speech, speaker verification, animal bioacoustics, and general audio on caller identification across four species and call-type classification across three. Differences between the strongest speech- and animal-pretrained representations range from $-0.027$ to $+0.119$ UAR; speech is significantly better in three of seven settings ($p<0.01$) and not significantly different in the remaining four. Controlled comparisons show no systematic advantage from increased language coverage or animal-domain pretraining, while frozen speaker-verification embeddings transfer poorly. The results also show that transfer is strongly layer-dependent: raw-waveform models peak early, whereas patch-spectrogram models peak deeper.

Regional accent cues can be captured under matched conditions, but it remains unclear whether they persist between read and spontaneous speech. We study RVG1, with 500 German speakers from nine regions, comparing ten speech representations on regional classification and continuous geolocation under matched conditions and speaker-independent read--spontaneous transfer. Whisper performs best under matched conditions, reaching 0.489 nine-way UAR and 148 km median geolocation error, but drops to 0.11/0.18 UAR across transfer directions and 363 km geolocation error. Self-supervised models show a similar degradation, whereas speaker embeddings are less discriminative in-domain but more robust under transfer. This contrast is consistent across classification and geolocation. Across representations, robustness is associated with how little a representation shifts between styles (style-invariance), for which crossstyle speaker retrieval is an interpretable proxy. Age, sex, sentence-overlap, and duration controls do not account for the gap, although channel characteristics contribute. These results show that strong matched-condition performance does not indicate robust regional information.

A dubbed character keeps its identity while the voice actor changes, so character identity and speaker identity are distinct properties of one recording, yet speaker verification measures only the latter. To address this gap, we propose KyaraBench, a benchmark that scores character voice directly instead of speaker voice, built from 85 human-audited identities in a dubbed anime corpus. It poses two challenging conditions: one that swaps the performer under a fixed character, and one that fixes the performer under changing characters. Listening studies with 78 participants provide human reference scores for cross-performer verification and same-actor discrimination. Speaker-verification baselines show increased errors under these character-specific conditions. We then train KyaraEmbed, a compact encoder using multilingual character supervision, same-actor negatives, and a language-alignment term. The model achieves the best performance on all character-specific conditions in the main comparison. We release the benchmark, protocol, and encoder publicly.

Speech-based Alzheimer's disease (AD) assessments increasingly rely on pretrained self-supervised learning (SSL) models that learn acoustic representations directly from raw audio, exposing the model to recording factors. We ask whether such factors are merely encoded in SSL representations or can systematically alter predictions. Using ADReSSo and three large SSL backbones, we apply controlled noise and reverberation interventions to participant-speech-only, non-speech, and full-recording audio. We combine layer-wise linear decoding, input- and representation-space interventions, and geometric alignment analysis to distinguish acoustic decodability from influence on AD prediction. Our results show that controlled acoustic interventions alter AD predictions across all three SSL backbones. Noise, despite showing no significant diagnostic-group difference in the original data, produces the strongest intervention effects. Importantly, these effects are systematically structured relative to the classifier's decision direction, replicate on the held-out test set and reverse when the representation-space intervention direction is reversed. Together, these findings show that high predictive performance and the absence of a significant diagnostic-group difference in a measured acoustic factor are not sufficient for robustness. We argue that intervention-based robustness tests should become standard for trustworthy clinical speech models.

Large Audio Language Models (LALMs) rely on effective audio encoders for multi-task performance. We introduce UniAE-MoE, a unified audio encoder designed to model cross-domain audio representations and achieve outstanding downstream understanding performance via a Mixture-of-Experts (MoE) architecture. Specifically, we explore mainstream audio encoders and integrate those from Qwen2-Audio and Audio-Flamingo 3, which demonstrate superior downstream capabilities. To facilitate effective model fusion, we improve our encoder using SwiGLU with shared experts to decouple encoder networks, and we further introduce a two-stage instruction-tuning strategy to better adapt the model to diverse downstream tasks. Moreover, we propose the task-specific data scaling (TSDS) technique to enhance UniAE-MoE's understanding capabilities. On the XARES-LLM benchmark, UniAE-MoE attains a score of 0.802, achieving state-of-the-art performance. It also delivers top-tier performance in the official Interspeech 2026 Audio Encoder Capability Challenge, further demonstrating robust generalization across diverse audio tasks. Together, these results validate the effectiveness of UniAE-MoE for unified audio understanding across speech, music, and general audio domains.

Real-time voice conversion (VC) systems commonly rely on pretrained speaker embeddings from automatic speaker verification (ASV) models. While effective for speaker discrimination, these embeddings are trained to remain stable across phonetic and prosodic variations within-speaker, which may conflict with frame-level acoustic generation in streaming constraints. To address this issue, we propose VOSSA (Voiceprint Optimization for Streaming Speech Architectures), a speaker representation framework that extracts speaker information from intermediate content encoder layers and aggregates using attentive statistics pooling. The embedding is trained jointly with VC objectives, removing the need for a separate speaker encoder. Across six datasets, VOSSA improves F0 dynamics and vowel-discriminative acoustic cues while maintaining comparable NISQA-MOS, WER, and speaker similarity. Perceptual tests further indicate improvements in naturalness, speaker similarity, intelligibility, and vibrancy.

Audio is increasingly used for human activity recognition (HAR) because it captures object interactions, environmental events, and contextual cues in everyday environments. High-resolution (HR) audio provides rich acoustic information for model development but incurs substantial energy and storage costs and may expose sensitive speech content. Low-resolution (LR) audio offers a more privacy-preserving and resource-efficient alternative for deployment, but reduced sampling rates can remove acoustic cues essential for activity recognition, leading to significant performance degradation. We formulate this training-deployment mismatch as sensor-resolution privileged learning, in which HR audio is available during training, while inference relies exclusively on LR audio. We propose RAST, a resolution-aware transfer framework that compresses HR teacher representations by preserving token-level information and neighborhood structure before performing localized HR-LR alignment. Experiments on the SAMoSA and AudioIMU datasets show that RAST consistently outperforms LR-only training and direct teacher-transfer baselines, improving LR-only recognition by up to approximately 7.8% while requiring only LR audio at inference.

Bandwidth extension, the task of reconstructing the high-frequency components of an audio signal from its low-passed counterpart, is a long-standing problem in audio processing. In this work, we extend recent advances in neural architectures by framing bandwidth extension as an audio token prediction problem. Specifically, we train a transformer-based language model on the discrete representations produced by a disentangled neural audio codec, where the disentanglement is guided by a Harmonic-Percussive decomposition of the input signals, highlighting spectral structures particularly relevant for bandwidth extension. Our approach introduces a novel codec design that explicitly accounts for the downstream token prediction task, enabling a more effective coupling between codec structure and transformer modeling. This joint design yields high-quality reconstructions of the original signal, as measured by both objective metrics and subjective evaluations. These results highlight the importance of aligning codec disentanglement and representation learning with the generative modeling stage, and demonstrate the potential of global, representation-aware design for advancing bandwidth extension.

Music producers search libraries of isolated instrument recordings, called stems, for sounds resembling parts of an existing song. Neural retrieval systems address this by mapping audio to embeddings and ranking library stems by their similarity to the query. The leading method, Contrastive Instrument Retrieval (CIR), encodes the mixture as a single embedding, but it works best when a user specifies the target's instrument family. We introduce Stembed, which encodes a mixture as several slot embeddings representing candidate stems. During training, we construct mixtures from stems of the same song and match their slot embeddings to those of the isolated stems. The slot embeddings from mixtures inherit the stem identities of their assigned solo embedding, enabling a contrastive loss. On mixtures from held out MoisesDB artists, Stembed outperforms a CIR-style baseline when both search the full stem library. Even when predicting the stem count itself without family labels, Stembed exceeds the baseline's family-filtered R@1. Our website demonstrates how users can select a slot by inspecting the tags of its retrieved stems.

Speaker verification (SV) models are commonly assumed to better capture nuances among speaker characteristics as verification accuracy improves, leading to their widespread use as automated proxies for human voice similarity in speech generation tasks. However, by establishing a human perceptual alignment metric and conducting systematic analysis, we demonstrate that perceptual alignment is governed far more by how a model is trained (its learning objective) than by how well it performs (EER). Notably, standard margin-based classification losses (e.g., AAM-Softmax) yield substantially lower perceptual alignment than prototypical metric losses, while EER itself fails to track human judgment, directly challenging the community's implicit assumption. We trace this divergence to embedding geometry, where a model's effective dimensionality ($d_{\mathrm{eff}}$) tracks perceptual alignment with a $-0.95$ rank correlation, revealing that the dimensional spread favored by classification losses fundamentally clashes with the low-dimensional nature of human voice perception. Imposing a dimensionality bottleneck compresses $d_{\mathrm{eff}}$ and raises perceptual alignment ($ρ_{\mathrm{align}}$) from 0.08 to 0.74, establishing a principled geometric criterion for evaluating voice similarity.

Symbolic models make melody, harmony, rhythm, and form explicit but typically stop before a finished recording; audio models produce complete songs while leaving composition implicit. We introduce YuE2, which unifies symbolic and audio music generation at frontier quality through symbolic planning. A single AR-NAR Mixture-of-Transformers (MoT) first writes a readable score specifying melody and harmony, expands it into semantic music tokens, and realizes it as full-song audio. In comparisons using the same checkpoint, experts prefer symbolic planning for overall quality and musicality, with 49.3% of overall preferences versus 34.6% without planning. Experts also favor the unified model over a separate language model and diffusion Transformer. On WildSongBench, YuE2 scores 6.73 on SongBench Global Avg, exceeding all evaluated public baselines. Selecting from eight candidates (best-of-8), YuE2 reaches 6.96, the highest observed mean among all evaluated systems. Expert listening further establishes its competitiveness with proprietary song generators, favoring best-of-8 over Suno v4.5 and yielding nearly balanced preferences against Suno v5. To learn this generation process from recordings without aligned scores, we introduce MERT2 and SheetSage2 to supply semantic and symbolic supervision. MERT2 sets a new state of the art in music representation learning, surpassing previous best results on 14 of 15 MARBLE metrics; SheetSage2 leads 12 of 15 benchmark-metric pairs in our lead-sheet transcription comparison. The same checkpoint follows score edits while largely preserving unedited musical content and generates zero-shot covers without cover-specific training. Its readable score also enables agentic music editing, with external language models translating user feedback into revisions of the composition.

The transition from vocoder-based to neural-codec speech synthesis makes generalization more difficult for speech deepfake detectors, particularly those relying on speech-oriented representations. It remains unclear which acoustic representations retain discriminative information when the generation mechanism changes. We therefore compare 12 acoustic representations using a shared low-capacity linear classifier to identify effective evidence under codec shift. The analysis shows that hierarchical XLS-R leads on the pooled test set, while pooled no-vocals residual statistics perform best on the unseen-codec condition, revealing complementary behavior across generation conditions. Building on this finding, we propose MN-P, a dual-view detector that integrates an utterance-level pooled no-vocals representation with token-level XLS-R features through adaptive gating. The proposed MN-P reduces EER by 54.2% overall and by 60.9% on the codec-unseen condition relative to the best-performing retrained state-of-the-art system, with consistent gains across different detector backends. These results indicate that pooled no-vocals residual statistics provide effective complementary evidence for cross-generation speech deepfake detection.

Continuous audio encoders compress audio along feature and time axes through latent width and frame rate, but their joint effect on downstream performance remains unclear. We train sixteen encoders spanning four widths and four frame rates, with downstream adapters and probes, using matched training protocols. Despite generally improved reconstruction at larger widths, automatic speech recognition (ASR) and spoken question answering (SQA) favor moderate widths at higher rates, with the best observed widths shifting toward larger values under stronger temporal compression. Frozen-model PCA interventions reveal distinct reconstruction and recognition sensitivities: removing the trailing half of the components substantially degrades ASR in high-rate 512-dimensional encoders with comparatively small reconstruction penalties, whereas 1024-dimensional encoders largely preserve both. Yet the projected 1024-dimensional model underperforms unmodified narrower models on ASR at 12.5Hz. These findings identify a width--rate interaction in downstream utility and suggest that how representations are organized during training matters beyond reconstruction fidelity and compressibility.

Text-to-audio (TTA) generation aims to synthesize realistic audio that faithfully reflects natural-language descriptions. Most TTA systems adopt a two-stage latent paradigm: an audio tokenizer is optimized for reconstruction and then frozen, after which a generative model is trained in the resulting latent space. However, reconstruction-oriented representations may be suboptimal for generation, motivating joint representation and generative learning. To this end, we introduce Unite-Audio, to our knowledge, is the first to jointly learn continuous audio representations and latent flow matching for TTA. By coupling reconstruction with self-supervised generative prediction, Unite-Audio allows the generative objective to directly shape the latent space rather than treating it as a fixed intermediate representation. We further employ Flow-GRPO post-training to improve text-conditioned generation. Experiments show competitive TTA performance with a compact latent flow model, while ablation studies confirm the benefit of jointly learning the audio representation and generative model. Audio samples are available at https://runwushi.github.io/Unite-Audio.

Speech technology penalizes some voices: recognition errs nearly twice as often for Black speakers, and accuracy declines for second-language accents and older speakers. We introduce TRIAD, an audit grid crossing 120 texts, 24 rendered demographic voice profiles (gender, age band, accent), and ten expressive styles via controllable text-to-speech, isolating perceived demographic attributes from content and affect. For ten open-weights encoders we define axis-fidelity functionals, principal-angle leakage between axis subspaces, and group-conditional gaps; a proposition proves that average probe disparity grows with the same aggregate voice-semantic leakage $Λ$ we measure, and a corollary shows that peak leakage forces worst-case disparity inside an active region. The measured mean-square probe disparity tracks $Λ$ (Pearson r = 0.93), and a black-box protocol exposes the same signature in two closed-source models. ORCA, an adapter combining axis-specific contrastive heads, an orthogonality penalty, and group-balanced sampling, cuts leakage 72% and roughly halves the gaps.

Compositionality, the ability to represent complex acoustic scenes as combinations of simpler sound sources, is central to auditory perception and classical additive signal models. Still, it remains unclear whether modern pre-trained audio representations internalize additive structure without compositional supervision. Existing evaluation frameworks of audio compositional reasoning largely focus on cross-modal audio-text alignment, leaving open whether audio representations themselves exhibit additive compositional structure independent of text grounding, analogous to vector arithmetic in word representations. To investigate this, we adopt a two-step diagnostic for frozen audio representations. First, we quantify linear alignment between representations and sound source labels using canonical correlation analysis. Second, we test additive compositional generalization via leave-one-combination-out reconstruction, grouping clips by exact source-label set, averaging their representations, and predicting held-out means from per-source contributions fitted only on training combinations. With larger combination holdouts, CLAP outperforms the permuted and label-overlap baselines on FSD50K, while the speech models do not outperform the label-overlap baseline. We examine representations generated by Wav2Vec2, HuBERT, and CLAP on FSD50K and CHiME-Home datasets. All three models show consistently higher linear correlation and more accurate leave-one-combination-out reconstructions than the permuted baselines. However, only CLAP shows large cosine similarity gains, which could be associated with its training on many kinds of audio and text. Finally, we note that all three models exhibit reconstruction residuals, revealing limits of additive compositionality such as nonlinear or non-compositional audio structure.

Large Audio Language Models (LALMs) utilize either continuous features or discrete tokens, yet the optimal representation paradigm for general audio understanding remains debated. Existing benchmarks often focus on narrow domains or evaluate encoders outside LALM contexts. To address these gaps, we systematically evaluate continuous and discrete representations across speech, sound and music. Utilizing our UniARC framework with dual evaluation strategies across model scales from SmolLM2-135M to Llama-3-8B, we analyze the dynamic relationships of data volume, model capacity, and computational efficiency. Our results reveal the pivotal role of semantic constraints in tokenization for audio understanding and demonstrate that scaling backbones fail to compensate for information loss in audio representation, especially in data-limited tasks. These findings offer practical guidance for balancing semantic density, fidelity, and efficiency in future LALMs.

Sample identification (SI) is the task of matching pairs of tracks, where one track is created by musically transforming an element of the other. In the absence of sample annotations at scale, the dominant training paradigm has depended on artificially creating sample pairs. Although a recently released dataset provides annotations of real sample pairs at scale, an effective training recipe is missing. In this work, we present SI Embeddings (SIE), an SI model that achieves state-of-the-art results on three benchmarks, including a large-scale test set. We show that the previous state of the art trained on artificial pairs generalizes only partially to real pairs, and that its training data limits its performance. We also show that real pairs do not fully account for SIE's performance: its architecture and training recipe contribute substantially. We provide the first fully supervised training recipe for real-world SI, establishing a strong foundation for future research in the field.

Self-supervised learning (SSL) has improved speech representations, yet performance degrades in pathological domains such as electrolaryngeal (EL) speech, and the computational footprint of SSL models limits their applicability in real-time, on-device deployment. We propose a multi-teacher knowledge distillation framework to train a lightweight, streaming content encoder that generalizes across healthy (HE) and EL speech. Two teachers are distilled progressively: a frozen SSL model providing discrete phonetic cluster targets from HE speech, and an EL-fine-tuned speech recognition model supplying continuous bottleneck feature targets. Evaluated via downstream speech recognition, our approach reduces the EL word error rate to 21.2%, compared to 39.3% for the strongest zero-shot SSL baseline. Among causal convolutional, Transformer, Conformer, and Mamba-based student architectures, a Mel-Conformer achieves the best combination of EL accuracy and computational efficiency. The final encoder contains 21.9,M parameters and runs at a real-time factor of 0.30 under ONNX Runtime on a single CPU core.

Few-shot audio classifiers may rely on foreground-background co-occurrences and fail when those correlations shift. On SpurAudio, the resulting representation shift is concentrated and class dependent: for ResNet12, the top 10 percent of channels explain 82.80 percent of the null-corrected shift contribution. We propose SAMPLESELECT, which predicts a fixed-budget feature mask independently for each input while keeping the encoder and source classifier frozen. Training uses differentiable Gumbel Top-k selection with foreground classification and cross-background contrastive losses; inference uses deterministic Top-k masks and support-only linear adaptation. Across ResNet12 and Conv64 in 5-way 1-shot and 5-shot evaluation, SAMPLESELECT gives the best OOD accuracy among the compared methods and improves the matched full-representation control by 4.90-8.38 percentage points. Ablations and representation analyses further support the learned selection mechanism. Code is available at https://github.com/Cross-Innovation-Lab/SAMPLESELECT/

This paper establishes a perfect-reconstruction (PR) interpretation of AddUNet and its full-rate realization, and introduces a Residual Full-Rate PR architecture for task-directed representation learning. The survivor--skip structure of a constrained additive U-Net is shown to be exactly equivalent to a critically sampled multirate PR filter bank. The full-rate formulation removes the complementary-subband restrictions of the critically sampled system while preserving PR. A Residual Full-Rate PR architecture is then proposed to progressively route task-irrelevant, nuisance, or redundant structure away from the task-facing survivor while retaining the routed information explicitly. Exact reconstruction is guaranteed for arbitrary shape-compatible linear or nonlinear routing operators, without requiring invertibility, a matched synthesis bank, reconstruction loss, or learned decoder. The resulting architecture decouples representation design from reconstruction design: conservation is structural, while learning is devoted to task-directed routing. The same formulation identifies an identity-shortcut ResNet with its residual output retained as a full-rate PR system. Experiments verify exact single-channel routing of linearly separable factors to machine precision. On TIMIT, the proposed front-end improves test PER from $28.60\pm2.09\%$ to $25.76\pm0.41\%$ with the recognizer and training protocol held fixed, while maintaining exact reconstruction. Speaker probing further shows that structural conservation does not itself imply task-specific invariance.

Procedural audio has emerged as a viable source for transferable audio representation learning, but its design principles remain unclear.We revisit two questions: how a procedural source should be scaled, and whether training choices developed on natural audio should transfer unchanged to procedural data.Using a controlled source, we separate scale into formula-class coverage C and within-class rendering diversity I.Experiments with FDSL and AudioMAE show that these two forms of scale provide different benefits and depend on the learning formulation and downstream task. A matched AudioMAE study further shows that procedural audio favors low mask ratios (10%--25%), whereas AudioSet-28K favors 50%--75%. Shared-codebook analysis reveals lower patch diversity and stronger temporal predictability in procedural audio. These results motivate source-aware procedural pre-training, where source scaling and learning configuration are considered jointly.Code is available at https://github.com/Cross-Innovation-Lab/Formula-Bank.

Universal audio representations must preserve acoustic detail while making high-level concepts accessible across speech, music, environmental sound, and downstream models of different capacities. We study semantic refinement of an acoustically pretrained encoder by adding audio-description alignment to a foundation of BEST-RQ, reconstruction, and CTC. We compare matched control, shuffled-description, and correctly paired trajectories to distinguish correct correspondence from an extra contrastive objective. Each endpoint is frozen and evaluated with a temporal-mean linear probe and a sequence-aware LLM readout, testing whether the refined information is directly accessible and remains useful to a stronger model. Across three paired seeds, correct alignment improves domain-balanced classification by 4.66 points with the linear probe and 2.59 points with the sequence-aware LLM, with positive changes in every domain. Correct pairing accounts for 87% of the linear-probe gain, while the LLM shows its clearest correspondence-specific benefit in captioning. Dense acoustic objectives provide complementary gains under both readouts. A separate 24-layer continuation remains competitive with leading public encoders under the shared evaluator, supporting the recipe beyond the controlled study.

System-wide Dictation on Apple devices runs entirely on-device, and the speech it transcribes reaches the foundation model through a tokenizer: an encoder that maps short windows of waveform onto the representation the language model reads. Because that model is sparsely activated under Instruction-Following Pruning, only a small subset of its experts occupies DRAM at any time, so the always-on tokenizer competes for the same memory, and its parameter count bears directly on power and latency. In this work we study how to compress such a tokenizer by distillation, taking as the supervision target neither the discrete token nor the output distribution but the pre-quantizer latent the model actually consumes - the last representation the two token interfaces share. We train only the student encoder to regress the teacher's per-frame latent under a squared-error objective, with a single affine layer absorbing the teacher-student width mismatch. Because the target precedes both the quantizer and the language-model bridge, one recipe covers both token interfaces we support, and applies both to a tokenizer pretrained alone and to one jointly trained with a language model. At 2.8x compression the distilled student stays within 1.9% relative WER of its teacher on five of six teacher-student pairs without any fine-tuning, and improves on an independently trained tokenizer of identical capacity by 3.9% relative.

Joint-Embedding Predictive Architecture (JEPA) has shown strong performance in learning rich representations through self-supervised prediction in latent space. However, it typically relies on teacher--student architecture with an EMA to stabilise training, and can tend to yield uninformative representations. Contrastive learning is stable to train and produces strong global representations, but remains limited on local tasks by the global nature of its objective. In this work, we combine both into CoJEPA: a single shared backbone jointly trained with a JEPA objective on masked sequence tokens and a contrastive objective on the class token. The contrastive gradient provides stability, removing the need for an EMA teacher entirely, while JEPA enriches the sequence tokens via local predictions that contrastive learning alone cannot provide. Crucially, no extra parameters are added to the backbone: the same model is guided towards richer representations purely through the design of its training signal. CoJEPA takes the best of both worlds, outperforming or matching both individual methods across global and local MIR tasks, with a particularly strong advantage on tonal and harmonic understanding, and without any task-specific architectural changes. CoJEPA shows that combining objectives with complementary inductive biases can substitute for scale, encouraging future work to invest in smarter training objectives over ever-larger models.

Computational analysis of music often relies on discrete representations, yet many musical traditions are organized around continuous pitch movement that resists segmentation into note-like units. For such traditions, the discrete units that analysis would build on are not given in advance. We address this gap by learning a vocabulary of local pitch-contour patterns directly from unlabeled audio, using a VQ-VAE that quantizes fixed-length contour segments into a finite codebook. To make the learned tokens stable across segmentation positions and small variations in timing and pitch range, we train the model with a reconstruction objective evaluated under the best alignment among a set of candidate temporal and pitch-domain transformations. Applied to Korean traditional music, the learned tokens recover information about expert-defined sigimsae categories without supervision, and in pansori individual tokens align with the two principal modes, Gyemyeonjo and Ujo, supporting their use as units for corpus-level analysis of contour-centric traditions.

We present DINO-A, an adaptation of self-distillation from vision to general audio representation learning. While DINO has become a canonical method in self-supervised vision and prior audio work has explored latent prediction (BYOL-A) and masked modeling (Audio-MAE, BEATs), no prior work has brought canonical DINO to general audio classification in the way BYOL-A brought BYOL. DINO-A retains DINO's multi-crop, EMA teacher, and high-dimensional projection, replacing only the input modality and augmentations with log-mel spectrograms and the BYOL-A v2 augmentation block. We pretrain three backbones, two Vision Transformers with 8x8 and 16x16 patches and a convolutional encoder, on FSD50K and evaluate them with linear probing on ESC-50, Speech Commands v2, UrbanSound8K, and GTZAN. Three findings characterize the resulting representations. Patch resolution within the Vision Transformer family has consistent effect on representation quality, with smaller patches winning across all four tasks. The choice between Vision Transformer and convolutional backbone interacts with task type: convolutional networks lead on speech while Vision Transformers lead on environmental sounds and music. Under identical pretraining and evaluation conditions, DINO-A and BYOL-A v2 differ by 11.96 percentage points on average, and we trace this difference to two mechanisms: the interaction between DINO's high-dimensional projection space and FSD50K's limited scale, and the additional cost of multi-crop augmentation, which DINO uses but BYOL-A v2 does not. The high-dimensional projection space, central to DINO's success in vision, becomes a liability at FSD50K scale.

Audio effects (Fx) representation learning plays a key role in intelligent music production, including automatic mixing and Fx style transfer. Existing methods typically rely on dry or nearly dry references for effect modeling, yet truly unprocessed audio is rarely available in practice, as real recordings inevitably reflect the microphone, room acoustics, and preceding signal processing. Instead of pursuing absolute effect encodings, we argue that the relative effect distance between audio signals is more meaningful for real-world music production. Motivated by this, we propose RelFx, a contrastive learning framework that learns relative effect transformations from general audio collections without requiring dry references during representation training. Our approach uses a dual-branch Siamese encoder equipped with cross-attention and differential gating fusion to infer the shared effect transformation from a reference clip and an effect-processed, content-related clip. We further propose an antisymmetric fusion variant for bidirectional effect encoding, such that swapping the input order directly produces a nearly sign-reversed embedding, a property not explored in earlier work. Moreover, our dry-reference-free formulation eliminates the reliance on dry multitrack datasets and enables training on effect-bearing audio. Experiments on Fx style transfer demonstrate state-of-the-art performance under the standard Fx-Encoder++ MUSDB18 evaluation protocol, consistently outperforming existing approaches across all four instrument categories.

Public infant cry corpora are small, label-incompatible, and almost always evaluated one corpus at a time. We ask what this practice hides and what fixes it. Across four cry corpora screened by a multi-level leakage audit (byte-level and embedding-level deduplication plus a within-corpus train-test near-duplicate audit), we probe four frozen encoders and a handcrafted baseline under a unified five-class need ontology and shared task formulations. The audit exposes what single-corpus evaluation conceals: within-domain macro-F1 swings by 0.57-0.80 for the same encoder, cross-corpus transfer is negative on average (negative-transfer ratio 0.19-0.35, significant in 18 of 30 directed cells, BH-FDR), and 349 content-identical clip groups carry conflicting metadata labels across corpus distributions. The same audit, however, reveals a consistent way forward. Transfer into the noisiest corpus is consistently positive in effect size at matched training size and after near-duplicate removal, offering a practical recipe for small, noisy corpora. Frozen probes saturate at modest label budgets, while stabilized fine-tuning wins with full labels; domain-adaptive pretraining significantly beats stabilized fine-tuning at 5-10-shot (the 1-shot advantage is not robust to optimization-seed variance) but shows no significant advantage at 50-shot or beyond. In the tested binary, shared-label settings, ontology-mapped joint training wins in all four encoder-by-target combinations, whereas naively merging unmapped labels costs up to 37 F1 points. We release the ontology, mapping code, and audit pipeline, turning incompatible cry corpora into a usable joint-training resource.

Residual connections are fundamental to deep speaker recogni- tion models, such as ECAPA-TDNN and ResNet. However, standard identity mapping limits information flow to a sin- gle path, constraining representation capacity. We introduce Manifold-Constrained Hyper-Connections (mHC), reformulat- ing residual paths as a multi-stream evolution where informa- tion is mixed through a doubly stochastic matrix. By employing Sinkhorn-Knopp iterations, mHC ensures energy conservation by preserving signal intensity and feature mean, which stabi- lizes gradients and mitigates signal degradation in complex net- works. We evaluate mHC by replacing standard residual con- nections in backbones including ECAPA-TDNN, ResNet-34, Res2Net, and E-Res2Net. Extensive experiments on VoxCeleb1 demonstrate that mHC connections consistently enhance per- formance across all architectures, highlighting its effectiveness for robust speaker representation learning.

Neural audio codecs serve as fundamental tokenizers for LLM-based audio generation. While semantic priors are widely exploited to enhance linguistic intelligibility, the integration of explicit acoustic priors remains underexplored, limiting synthesis fidelity in frequency-sensitive domains. To address this gap, we introduce MeloCodec, a novel framework designed to effectively incorporate melodic priors, a critical form of acoustic information for singing. To address the optimization instability typically caused by the direct fusion of such explicit priors, we propose a Tokenize-then-Fuse paradigm that pre-trains a discrete melodic branch to lock in structures before feature fusion. To robustly realize this paradigm, we further propose a two-stage training strategy that prevents codebook collapse and ensures stable convergence. Experiments show that MeloCodec outperforms baselines in singing voice representation, improving pitch consistency and enabling controllable pitch manipulation with minimal timbre degradation.

Scaling the corpus is the default remedy when a contrastive representation lacks an attribute. We report a case where it does nothing, and identify what does: adding a lexical-speech round to a frozen-base multimodal embedding model raises zero-shot keyword spotting by 76 points while reducing speech-emotion recognition by 14. The loss is not a capacity limit: fine-tuning on 7,442 clips from a prosody-controlled corpus recovers emotion past its pre-speech level at a five-point keyword cost. Nor is it data volume: 29,428 mined clips whose captions explicitly name emotions, at matched exposure, move emotion by -0.0007. The difference is structural: a contrastive objective encodes an attribute only when the in-batch negatives cannot be separated without it; the controlled corpus holds sentence content fixed, so prosody is the only separating signal, whereas mined captions name emotion yet remain separable by scene content. Intervention on the same audio confirms causality: raising caption similarity does not recover emotion, but collapsing caption diversity so that emotion becomes the only separating axis recovers it by 8.9 points across three seeds, with a smaller, same-signed gain on a non-acted corpus, while keyword accuracy trades back. Corpus structure, not size or caption vocabulary, controls what a contrastive audio embedding encodes.

Balancing sequence length, representational capacity, and long-horizon stability is a central problem in autoregressive (AR) speech and audio generation. Representations with higher frame rates or greater capacity can preserve more signal detail, but they also make streaming generation more vulnerable to distribution drift and AR error accumulation. Conversely, shorter and more compressed representations simplify AR modeling, but their limited bandwidth may discard important components and constrain the upper bound of reconstruction fidelity and generation quality. We ask whether a low-frame-rate, high-dimensional, high-bandwidth continuous representation can be co-designed with a streaming generation framework to support robust high-fidelity reconstruction, strong single-token predictability, and superior long-horizon stability. We decompose this goal into two coupled problems: what geometric and statistical properties a high-dimensional representation space should have, and how an AR continuous-token generator should be structured to resist error accumulation. Accordingly, we propose Locodec, a locally encoded codec that shapes its representation space to improve the interpolatability of a lower-dimensional core manifold and the identifiability of the native high-dimensional coordinates, thereby improving the predictability of high-dimensional high-bandwidth tokens. We also propose MP-ELD, a single-token AR flow-matching framework that uses multi-path information routing and residual classifier-free guidance to mitigate error accumulation. Experiments with 8-Hz, 768-dimensional tokens show that our design preserves reconstruction quality, improves single-token predictability, achieves competitive WER, and maintains stable long-form synthesis, without using external SSL/ASR models, pretrained text language models, or post-training stages.

This study analyzes the intermediate representations of music foundation models (MFMs) and reports the geometric structures used to represent pitch information. By inputting isolated musical notes into trained MFMs and analyzing their principal components, we reveal that the representations form a helical structure reflecting the octave periodicity of pitch. Furthermore, we show that the clarity and geometry of this helical structure vary not only across models but also with the acoustic properties of the input signals. Our analysis provides a novel approach for clarifying the internal mechanisms of MFMs.

Neural audio codecs (NACs) have become popular for obtaining speech representations as discrete tokens. Beyond compression, discrete tokens can be used to train self-supervised learning (SSL) models. Such models, referred to as codec-based SSL models, reduce data storage and computational cost, enabling scalable SSL pre-training. However, their language sensitivity remains unclear. When the language changes, codec-based SSL models may require retraining, which undermines their efficiency. In this paper, we present a systematic analysis of language sensitivity by varying either the NAC training language or the SSL pre-training language while keeping the other fixed. Experimental results show that downstream performance is insensitive to the NAC training language but strongly dependent on the SSL pre-training language. These findings suggest that a single NAC can be reused across languages, while aligning the SSL pre-training language with the target language is crucial.

Latent Acoustic Mapping (LAM) is a self-supervised learning method that generates high-resolution spherical acoustic maps from multichannel recordings without labelled data, matching supervised baselines on direction-of-arrival benchmarks. However, LAM degrades significantly with sparse 4-channel arrays, as the low-resolution cross-spectral matrix captures far less spatial information than the 32-channel inputs LAM was designed for. We benchmark a diverse set of upsampling architectures, spanning lightweight convolutional networks, iterative back-projection models, physics-informed networks, and generative adversarial approaches. We also study whether aligning these upsamplers with LAM by training them jointly or in different stages helps preserve the spatial structure that LAM depends on. Results show that the original full-resolution LAM is the strongest, that separately trained lightweight models are the most competitive learned approaches, and that representation alignment between the upsampler and LAM matters more than model complexity.

Do learned audio embeddings encode structure that nobody told them to encode? We probe four large pretrained audio models (AST, CLAP, BEATs-bio and BirdNET) with a downstream task none of them saw during training: recovering phylogenetic distance from species vocalizations. If the geometry of the embedding space tracks the tree of life, the representation is picking up something deeper than the labels the model was optimized for. We run Mantel tests across two independent radiations. In 32 marine mammal species (1,754 recordings from the Watkins Marine Mammal Sound Database) the foundation models recover strong phylogenetic signal within the 26 cetaceans (CLAP r=0.82, BEATs-bio r=0.82, AST r=0.74; all p<0.001), among the highest acoustic-phylogenetic correlations reported for any taxon. Hand-crafted MFCC features (105d) find nothing (r=0.040, p=0.338). The gap survives after PCA-projecting every embedding down to 105 dimensions, so it is not an artefact of representation size. It also survives a partial Mantel test controlling for dominant frequency (partial Mantel r=0.404, keeping 97% of the variance explained), so it is not just pitch in disguise. We repeat the analysis on 20 bird species using the Jetz et al. (2012) phylogeny, and this time add BirdNET, a classifier trained end-to-end on around 6,000 bird species. The general-purpose foundation models recover the signal again (AST r=0.55, CLAP r=0.52). The unexpected result is that neither BirdNET nor the bioacoustic BEATs-bio beat them (r around 0.32 to 0.36). Matching the training domain to the target taxon does not, by itself, help. Pretrained audio embeddings carry evolutionary information across two independent radiations, and domain-specific pretraining is not required for it to emerge.

Sample retrieval tools can help composers find harmonically compatible material, but querying from a fixed reference sample becomes less informative as arrangements evolve and the harmonic context shifts with each musical decision. We present Reflector, an interactive audio workstation that tracks harmonic combinations as they accumulate on the composer's timeline and adapts retrieval as the arrangement develops. The system is organized around a fixed interval-class oracle: a hand-designed table of weights that scores how pitch-class content combines between sources. An encoder trained entirely on synthetic audio learns to approximate the oracle in a 128-dimensional embedding space, where dot products stand in for compatibility scores at interactive speed. As the composer arranges material on a multi-track timeline, a sweep-line analysis discovers co-sounding regions, computes oracle-weighted centroids, and retrieves against the composite harmonic identity of the session as it evolves. Session centroids projected into a navigable 3-D space reveal structural harmonic relations across the composer's body of work. This paper is a systems account: we give the design rationale for each architectural decision, characterize Reflector's behavior through intrinsic measurements on a working sample library, and describe the implementation. The characterization yields a central finding: the learned embedding preserves the kernel's pairwise judgments while covering the whole library, something the kernel cannot do when used directly as a retrieval rule, because the embedding's normalized geometry cannot express the degenerate solutions that direct scoring favors. The entire pipeline runs locally with no copyrighted training data. Reflector is free, and the training pipeline is open source.

Joint Embedding Predictive Architectures (JEPA) have recently emerged as a paradigm for learning world models by predicting latent representations, offering a promising direction for self-supervised learning. While initial attempts have applied JEPA to the music domain, it remains unclear how such frameworks can naturally support the formation of a world model for music. In this work, we propose to learn a world model of piano sound using JEPA by framing music as an action-conditioned system: the audio is treated as the state, and the pianoroll as the instrument action. Given a current audio state and an action, the model predicts the resulting future audio state, mirroring how humans learn musical sound through interaction. The model is trained in a fully offline setting using paired audio-pianoroll data, without environment interaction. Experiments show that the learned model captures the relationships between musical actions and their resulting sound. The resulting representations support downstream tasks, including beat tracking, composer identification, and key estimation, and enable piano transcription via planning, by searching for actions that best explain a target sound.

Neural audio codecs with residual vector quantization (RVQ) normally treat all frequencies uniformly, so their codebooks become spectrally entangled. Truncating stages then removes an unpredictable mix of frequencies. Parallel band decomposition addresses this by splitting audio into independent bands, but fragments the latent space and loses cross-frequency coherence. We introduce HARP (Harmonic-Aware Residual Partitioning), a training strategy that partitions RVQ stages into frequency-ordered groups where each group refines its target band while the decoder retains access to all lower frequencies. Overtones are reconstructed in the context of their fundamentals, preserving coherence that parallel methods lose. HARP requires no architectural changes; it only modifies the training loss, leaving inference identical to standard RVQ. On speech, music, and general audio, HARP outperforms both standard RVQ and parallel decomposition. MUSHRA listening tests also show perceptual improvements.

Audio mixing style encompasses the artistic and technical decisions a mix engineer makes, including level balancing, spatialization, and the choice, ordering, and parameterization of audio effects (FX) on each stem. FX chains are a key determinant of this style, yet existing approaches to modeling them remain limited. Some operate on stereo mixtures without explicit per-stem FX chain modeling, others fix the number or type of effects per track, and many require differentiable effect implementations or scarce multitrack datasets. We present StemFX, a framework that learns mixing style representations by autoregressively predicting variable-length FX chains on source-separated stems. A Transformer decoder predicts tokenized FX chains autoregressively, while a band-split multi-band CNN encoder with FiLM conditioning captures per-stem spectral structure. To enable large-scale paired training, we extract pseudo-stems from about 105K songs via source separation and augment them using MultiAFx, a toolkit unifying 85 audio effects from 7 Python libraries. Evaluated on mixing style retrieval, StemFX outperforms all baseline models across all tested chain lengths. On paired mixing style transfer, StemFX achieves the best spectral fidelity and the highest listener preference, over 4000 times faster than iterative optimization.

This paper details the DS@GT ARC team's approach to BirdCLEF+ 2026, multi-label detection of animal vocalizations in soundscapes from the Pantanal wetlands. The 2026 edition adds about an hour of labeled soundscapes, shifting the task toward supervised pipelines fit to the labeled set. First, we build a competitive supervised baseline that ensembles a frozen Perch v2 backbone, a trained HGNetV2-B0 sound-event-detection network, and a non-bird prototypical head, reaching a private leaderboard score of 0.936 at rank 1894 within a 90-minute CPU budget. Second, we ask whether token-based representations can compete, contrasting codec representations from neural audio codecs against semantic representations from foundational embeddings. We compare two bioacoustic specialist models against four token-based encoders trained on AudioSet. The repository for this work can be found at https://github.com/dsgt-arc/birdclef-2026.

Representation alignment (REPA) has been investigated to accelerate diffusion training, but we observe that regularizing intermediate representations in diffusion Transformers (DiT) may implicitly entangle latents and limit generative capacity. To address this issue, we propose ReGen, a hierarchical multi-prompt representation generation framework that jointly estimates multiple vector fields for both representations and data within a single diffusion model. We further introduce generalized flow matching (GFM) to improve the generalization of conditional flow matching (CFM). We validate ReGen on single-stage waveform diffusion models including neural audio codec and Wave-VAE. ReGen significantly improves waveform generation quality from highly compressed latent representations at 12.5 Hz. We also present ReGenVoice, a latent diffusion model (LDM)-based text-to-speech model that achieves strong speech intelligibility (WER) and speaker similarity (SIM) with a small dataset. Moreover, operating the LDM at 6.25 Hz with rich semantic and acoustic latent representation enables efficient training and sampling, requiring only 1 day of training on 4 GPUs and fast inference with an RTF of 0.08. Audio samples are available at https://regenvoice.github.io/demo/.

End-to-end neural audio models achieve high-fidelity compression and generation. We might read that performance as evidence they directly represent interpretable features such as pitch and timbre, but a model can produce plausible outputs without doing so. A model may encode these features in any reachable basis, but regardless of which, the features are well described as compositions of time-frequency-localized primitives. Whether state-of-the-art encoders preserve access to these primitives, and thus to compositions of them, remains unclear. Through theoretical analysis and controlled experiments, we show that several state-of-the-art strided convolutional encoders impose two structural bottlenecks, both predictable from architecture and signal structure, on access to these primitives: (1) they collapse primitives into alias equivalence classes, establishing a bound on representational capacity, and (2) they limit the frequency resolution available to learned filters, restricting separability. For well structured data, we find collapse rates of 31-35% and filter bandwidths 10-35x above the theoretical resolution bound, confirming that both bottlenecks arise under realistic signal conditions. We then introduce Gabor Latent Refactorization (GLRF), a lightweight post-hoc intervention that re-expresses encoder latents in a frequency-localized basis, reducing filter bandwidths from 10-35x to 1.5-3x of the theoretical resolution bound while preserving reconstruction fidelity and improving control over attributes like pitch. These results show that the encoders in question predictably degrade access to frequency-localized primitives, entangling the features that depend on them, and that a lightweight, retraining-free intervention can recover much of that access, improving steerability and interpretability.

Audio-language models compress a speech encoder's output through a Querying Transformer (Q-Former) connector before feeding it to a large language model. We identify two failures in this compression. The connector's output vectors collapse to a single direction, and different speakers produce nearly indistinguishable outputs, with paralinguistic cues such as speaker identity, gender, and prosody lost along the way. Our method, ORCA, reverses this collapse by splitting the queries into groups whose outputs are constrained to point in different directions. On SAKURA multi-hop reasoning, ORCA gains 26.4 points over an identically trained 4B baseline, reaching 75.2% (vs. 49.0% for the 8B Audio Flamingo-3). At the connector level, the same change cuts query redundancy by 12x and raises cross-speaker variance by 75x.

Passive hydroacoustic monitoring often generates large volumes of continuous recordings that are only partially exploited due to the cost of manual annotation. Supervised detection methods perform well but require large labeled datasets, seldom available for rare signals or understudied environments. This work proposes a self-supervised exploration pipeline to address this limitation in low-frequency settings. A Masked AutoEncoder (MAE) is pre-trained on a reconstruction pretext task, then used to extract patch-level representations from spectrograms. Within each spectrogram, adjacent informative patches are aggregated into event-level embeddings, enabling the disentanglement of overlapping events. These embeddings are then clustered at the dataset scale using the dimension reduction algorithm UMAP and the clustering algorithm HDBSCAN to identify hydroacoustic patterns. The pipeline was applied to a multi-year hydroacoustic dataset collected near Mayotte Island, Indian Ocean, containing marine mammal vocalizations, seismo-volcanic signals, and anthropogenic noise. The 317 clusters were manually mapped to 15 hydroacoustic classes or noise in less than one hour. The method was evaluated in two ways. Quantitatively, when used as a classifier, it achieved performance comparable to two existing detectors. Qualitatively, it recovered known seasonal patterns of marine mammal acoustic activity. It also identified patterns of previously unstudied signals, thereby demonstrating its practical value.

Neural speech codecs are increasingly used as intermediate representations in codec-based speech generation systems. TiCodec introduces a factorized representation that separates time-varying speech content from time-invariant information through a Time-Invariant Representation Extraction (TIRE) module, potentially reducing the amount of information that must be modeled at the frame-level. In this work, we investigate the nature of the information captured by TIRE representations and their suitability for low-latency speech processing. Using a series of probing tasks, we analyze the influence of the encoder layer and show that intermediate layers capture complementary speaker- and environment-related information while containing little linguistic content. We further study several segment selection strategies for TIRE training and demonstrate that cross-file sampling improves the robustness of invariant representations. Based on these findings, we propose Dual-TIRE, a multi-level architecture that exploits the complementarity of different encoder layers and improves speech reconstruction quality and speaker similarity. Finally, we evaluate TiCodec in a streaming inference setting using successive 660ms processing blocks. Results show that streaming operation can be achieved without significant degradation in reconstruction performance, highlighting the potential of factorized neural codec representations for future low-latency speech generation systems.

Modern automated audio captioning systems pair a frozen audio encoder with a large language model (LLM) via a trainable projector, incurring the encoder's inference cost and bottlenecking the model through its fixed acoustic features. We present CARD, an encoder-free audio captioning model that removes the encoder at inference: a 13.2M projector feeds a frozen LLM with merged LoRA adapters, while the teacher used to train it is discarded. CARD distills a pre-trained audio teacher (CLAP-HTSAT) into the model, but rather than injecting it into the LLM alone, it routes the teacher's representations across components: perceptual stages to the projector and semantic stages to the LLM. This placement improves CIDEr-D by +11.9 over an LLM-only distilled model on AudioCaps and by +5.0 on Clotho, reaching 55.4 against a 66.4 encoder-kept upper bound with no encoder at inference, showing that where a teacher's knowledge is placed matters as much as its presence.

Audio-conditioned generators now produce synthetic sound effects from real recordings, so the real and synthetic versions of an event increasingly coexist in sound libraries and in the corpora used to train audio models -- yet no benchmark measures whether a representation can match a synthetic clip to the specific real recording it was generated from. I introduce Doppelganger, a benchmark for matching sound effects across the synthetic-real boundary, pairing 10,420 real clips across 34 everyday sound events each with an audio-conditioned synthetic twin, alongside a controlled 7-class corpus. Off-the-shelf audio encoders do not cross the boundary cleanly. Making the embedding ignore the boundary the standard way -- training it on sound-event labels -- works on familiar sounds but backfires on new ones, dropping below the untrained encoder. Training on the pairs instead -- a clip and its own synthetic twin -- generalizes. On sound events held out of training, it recovers the exact real source about 80% of the time (up from 61% untrained; chance 0.03%), whereas no objective meaningfully improves category-level recognition on those unseen events. The learned matching is specific to one generator -- it survives changes to that generator's settings but not a switch to a different generator, and collapses for the text-only generators tested. A human annotation baseline (49 listeners) lands well above chance but below the models on the same trials. Synthetic twins fool people into calling them real about 29% of the time, yet a generator-specific detector separates these audio-conditioned twins from real recordings perfectly.

Audio foundation models are widely adopted as general-purpose feature extractors, yet the internal structure of their learned representations remains insufficiently understood. In this work, we analyze CLAP audio embeddings through a probing framework, studying the encoding of three fundamental perceptual dimensions: reverberation (RT60), loudness (LUFS), and spectral content, measured via spectral centroid (SC) and relative pitch (RP). Probes of increasing complexity are trained to predict each attribute from frozen embeddings across five datasets spanning noise, speech, monophonic musical notes, and music mixtures. Our primary finding is that all of these attributes are reliably recoverable from the CLAP embedding space across the examined datasets. Within this global picture, two encoding regimes emerge: RT60, LUFS, and RP are approximately linearly encoded, while SC requires non-linear probes. Both regimes generalize across eight additional audio foundation models, with the notable exception that amplitude-invariant architectures discard loudness entirely by construction. The identified linear feature directions are geometrically consistent across datasets for RT60 and LUFS, while highly domain-specific for RP. Finally, we provide a qualitative demonstration of cross-modal consistency, showing that text embeddings of acoustic descriptors align geometrically with the identified RT60 feature direction.

Crossmodal correspondences between sound and taste are well established in psychology and neuroscience, but largely absent from content-based multimedia retrieval. We formalise taste-from-audio prediction as a content-based music information retrieval benchmark over a perceptually validated multi-source corpus, comparing ten frozen audio encoders from the four HEAR families under a shared multi-task regression head, with gated late-fusion as a configurable variant. In order to assess the effectiveness of the models, we compute absolute error and rank correlation. The strongest systems predict the five tastes within a macro RMSE of 0.134; on held-out real music their error is less than half a single rater's deviation from the consensus (RMSE 0.13 vs. 0.28), so the model tracks the group consensus more closely than an average human rater, and well below the previous state of the art baseline (0.219). On absolute error the encoders are statistically flat, with a single VGGish matching the best fusion, but gated late-fusion's advantage is confined to rank correlation (macro Pearson r 0.724 vs. 0.666). Operationalised as a content-based retrieval index, the predicted taste space ranks a 309-item pool far more faithfully than a CLAP-text baseline, which sits at chance; ridge probes and an audio-bandstop knockout read the strongest representations against documented sound-taste correspondences.

Room embeddings derived from reverberant speech are often unreliable: speech content and recording degradation can alter the representation even when speaker, room, and source-receiver geometry remain unchanged, degrading downstream task performance. We propose a framework that learns room embeddings robust to speech-content variation and a representation-level uncertainty score from reverberant speech without downstream-task supervision. The embedding is anchored to a structured room impulse response (RIR) latent space and trained using a multi-view data structure with Kullback-Leibler (KL)-based alignment; a multi-positive contrastive term further refines robustness. A lightweight uncertainty head is calibrated using the dispersion of corruption-induced embeddings and optimized with a rank-based objective. Across waveform- and spectrogram-level corruptions, the score is consistent with representation dispersion and enables effective selective prediction while requiring only a single utterance at inference.

Recognizing jazz standards from audio is a challenging form of tune-level music retrieval: different performances of the same standard may vary in tempo, key, arrangement, instrumentation, improvisational content, and even whether the head melody is present. We study this problem using a curated subset of the Jazz Trio Database designed for cross-performance standard recognition. We compare a from-scratch trained Harmonic CNN baseline against frozen pretrained music representations from recent music understanding foundation models, using both supervised probing and nearest-neighbor retrieval. Our results suggest that from-scratch spectrogram models overfit strongly to training performances, while pretrained embeddings provide better top-$k$ results but are sensitive to performer identity, which can be partially reduced with a lightweight contrastive projection. Our findings motivate jazz standard recognition as a useful stress test for music representation models and as a step toward retrieval-based standard identification. Project page: https://github.com/cagries/tipofmyear.

Self-supervised learning enables audio representations that transfer across domains and tasks. We present BEST-RQ-2, an evolution of BEST-RQ that retains frozen randomprojection-based discrete targets while introducing a two-step contextualize-then-predict pretraining scheme. A ViT context encoder processes only the unmasked spectrogram regions, and a lightweight predictor infers targets for the masked regions; the predictor is discarded after pretraining. Replacing the original Conformer encoder with a ViT shifts performance across domains, slightly reducing speech performance while improving music and environmental sounds, with comparable average scores. The main improvement comes from decomposing masked prediction into separate contextualization and prediction stages. On the X-ARES and XARES-LLM benchmarks, BEST-RQ-2 consistently outperforms one-stage baselines in overall transfer while keeping inference compute unchanged. Code and model checkpoints are publicly available.

Recent advances in language--audio retrieval have been largely driven by contrastive dual-encoder architectures that align audio and text in a shared embedding space. While effective, existing retrieval embeddings are primarily optimized for audio--caption matching, limiting their ability to support diverse retrieval objectives and controllable retrieval behaviors. We present ALM2Vec, a universal audio embedding framework derived from pretrained large audio--language models (LALMs). By transferring the audio understanding, instruction-following, and reasoning capabilities acquired through large-scale multimodal training, ALM2Vec learns a unified embedding space for retrieval across audio domains and task types. Beyond conventional text--audio retrieval, ALM2Vec incorporates natural-language instructions into the embedding process, enabling instruction-aware retrieval for scenarios such as audio question answering and aspect-conditioned retrieval. Experimental results show that ALM2Vec achieves competitive performance on standard audio and speech retrieval benchmarks while exhibiting promising compositional and controllable retrieval capabilities, highlighting its potential as a unified audio embedding model for retrieval across domains, tasks, and user intents.

This report investigates the extension of pretrained General-Purpose Audio Tagging (GP-AT) models toward spatially grounded Sound Event Localization and Detection (SELD). The proposed AT2SELD framework couples a pretrained AT backbone with compact First-Order Ambisonics (FOA) spatial processing, track-wise SED and Cartesian DOA estimation, permutation aware supervision, and calibration. It characterizes how semantic audio priors support localization-aware scene analysis under data, computation, and deployment constraints. The framework is developed through informed multi-stage Neural Architecture Search (NAS). Stage 1 shows that spectral FOA descriptors, based on magnitude, phase, and Intensity Vectors (IVs), provide the most reliable interface for semantic-to-spatial transfer. Stage 2 identifies early residual spatial encoding as the main capacity-sensitive component, while late track-wise abstraction and recurrent smoothing act mainly as refinement stages. Stage 3 shows that late cross-stitch coupling improves semantic-spatial interaction, whereas early fusion is costlier and less effective. Diagnostic evaluation analyzes the selected architecture under class balancing, focal loss, activity-conditioned DOA supervision, threshold calibration, and transfer across STARSS23, TAU2019, TAU-NIGENS2020, and TAU-NIGENS2021. Focal loss improves the activity point, active-only DOA supervision mitigates inactive target dominance, and validation-selected thresholds recover calibration without replacing spatial learning. Cross-dataset and oracle-activity analyses indicate strong fixed source localization on TAU2019, transferable representations from TAU NIGENS2021, and meaningful but uncertain behavior on STARSS23. Overall, GP-AT priors appear promising for SELD design when embedded in spatial-aware architectures and optimized through integrated calibration and deployment oriented strategies.

Learning discrete speech representations that preserve similarity across variable-length utterances is central to query-by-example spoken term detection (QbE-STD). While wav2tok introduced CTC-based sequence alignment to enforce token consistency, its tightly coupled clustering and alignment training recipe limits scalability. We propose wav2tok 2.0, a scalable alignment-aware speech tokenizer built on the BEST-STD backbone. wav2tok 2.0 employs staged training, first learning discriminative, speaker-invariant representations via contrastive learning and vector quantization, and then enforcing pairwise token consistency using a CTC alignment loss and a novel DTW-aligned framewise prediction objective with adaptive weighting. Experiments show that wav2tok 2.0 consistently outperforms BEST-STD and general-purpose tokenizers on QbE-STD while remaining efficient and scalable.

While pre-trained models excel in specialized tasks, learning universal representations across diverse acoustic domains remains challenging. To address this, we propose WQ-Fusion, a robust dual-encoder framework for cross-domain audio representation learning. Overcoming the limitations of static concatenation, WQ-Fusion integrates whisper and qwen via an Adaptive Feature Modulation module and a novel element-wise gated attention mechanism. This design enables dynamic feature selection, allowing the model to selectively emphasize relevant acoustic and semantic dimensions. Extensive experiments on the Interspeech 2026 Audio Encoder Capability Challenge (Track A) benchmark demonstrate that by effectively routing heterogeneous information, WQ-Fusion achieves a superior overall score of 0.836, significantly outperforming the strongest single-encoder baseline.