Papers

Large audio language models answer questions about speech, sound, and music, yet their accuracy drops sharply on tasks that need fine-grained acoustic analysis. Judging which of two speakers has the higher pitch demands iterative signal-level reasoning that a content question does not. Current models spend the same computational depth on both. We introduce AdaLoop, a lightweight recurrent module that learns how many latent refinement steps a given audio–question pair requires. A shared transformer block iterates over the audio representation, guided by the question, while a learned halting mechanism exits the loop once the representation is ready. AdaLoop adds fewer than 3% of the base model's parameters and plugs into any audio encoder–language model pair without modifying either component. Evaluated on three architecturally distinct models across MMSU, MMAU-Pro, and MMAR, AdaLoop raises the average accuracy by 2.9 to 3.8 points, with the largest gains on perception-heavy subtasks where the model learns to apply deeper reasoning.

In recent years, flow-matching models have produced significant improvements in zero-shot text-to-speech synthesis. Conditioned on an audio prompt and text, these models learn a velocity field and generate speech by iteratively solving an ODE. During inference, the solver evolves a single state spanning both the prompt and the region to be generated, although only the generated region is ultimately retained. The discarded prompt state, however, still matters; its intermediate values influence generation through the velocity field that couples the two regions. As sampling continues, this state can drift away from the prescribed conditional path, introducing a discrepancy into subsequent generation updates. Unlike the unknown generated trajectory, the prompt path is available in closed form from the reference audio and the initial noise. We exploit this observation with Prompt-Consistency Inference (PCI), a training-free rule that restores the prompt block to its analytic value before each velocity evaluation, while leaving the generated block unchanged. PCI improves speaker similarity and intelligibility across the evaluated flow-matching TTS backbones without additional network evaluations. Our ablation studies further show that PCI keeps post-step prompt discrepancies smaller and that corrections covering the later sampling stages recover much of the observed similarity gain.

Neural audio codecs compress waveforms into compact discrete tokens that underpin speech language models, real-time communication, and large-scale audio storage. Almost every dominant design, including residual vector quantization, finite scalar quantization, and single-codebook variants, follows the VQ-VAE template by partitioning the encoder latent through a learned codebook or a fixed scalar grid. We ask whether this partition is necessary. We introduce GS-Codec, a neural speech codec whose bottleneck is a parametric signal decomposition rather than a quantizer. We adapt Gaussian splatting from 3D scene reconstruction to one-dimensional latents. An inner optimization loop fits each encoder segment as a weighted sum of 1D Gaussian primitives. The decoder then reconstructs the waveform from the rendered sum. To avoid the cost of this iterative inner loop at inference time, we additionally train a lightweight GS Predictor Net that regresses the primitive parameters in a single forward pass. The encoder and decoder are trained end-to-end through the inner loop, with no quantizer anywhere in the training pipeline: the bottleneck is the decomposition itself, and scalar quantization is applied only post-training to the fitted parameters. Rather than relying on discrete codebook stages for bitrate control, our representation exposes a fine-grained rate-quality tradeoff: a single trained checkpoint supports post-training bitrate control by varying the number of primitives and the per-parameter bit depth, with no retraining required. GS-Codec matches or exceeds well-established open-source codecs such as EnCodec and DAC on speaker similarity (SIM), intelligibility (STOI), and perceptual quality (UTMOS) at comparable bitrates, while achieving comparable semantic performance (WER). Code and audio samples are available at https://ronaluf.github.io/gs-codec/

In this article, we present a method of spectral serialization envisioned by G. Mahler and A. Schoenberg between the late 19th and early 20th centuries that enables the selection of timbres rooted in musical structure. Based on information geometry, we evaluate the Wasserstein distance between polyphonic timbre spectra, as demonstrated in our previous paper, in terms of the second moment of spectral chords comprising three or more notes, expressed in squared hertz. By embedding these quantities in a Banach space, we analyze specific passages from works by R. Strauss, Mahler, A. Berg, and A. Webern from the perspective of timbre, showing how the same chord exhibits different affinities and divergences depending on orchestration. Taking the wah wah mute as an example, we propose a model that treats instrumental timbre and speech within the same Banach space, and we attempt to quantify the "vowel triangle" in Hz. These efforts constitute a partial but integrated solution to the historical three problems of Schoenberg. Serialism, Klangfarbenmelodie, and Sprechgesang which neither Mahler nor Schoenberg in their lifetimes, nor the postwar generation of composers including Nono, Boulez, and Stockhausen, was able to resolve fully. We aim to reconsider musical thought as a whole from the perspective of timbre and to open new possibilities for interpretation, performance, and composition. All tools are publicly available on GitHub, so that anyone can reproduce and extend these analyses.

Adding expressive and environmental controls to a speech generator requires learning heterogeneous attributes without losing earlier capabilities. VoiceWeaver addresses this problem with structured label prefixes and staged emotion-tone-event training in a shared text-audio model. Replay-based distillation retains earlier predictions, while embedding decorrelation and attribute dropout regularize conditioning. Evaluation separates single-attribute correctness, joint emotion-event generation, and three-attribute correctness. Rounded to whole percentage points, emotion accuracy is 86% in Chinese and 83% in English. Chinese emotion-event joint accuracy is about 70%, versus 56% for mixed-task training and 48% for Ming-omni-tts; English joint accuracy is 66%. Three-attribute accuracy is 59% in Chinese and 55% in English on 500 samples each (57% pooled). Text error rates exceed those of the external TTS baselines. Audio samples are available at https://anonymous.4open.science/api/repo/VoiceWeaver1-4D88/file/index.html.

Although reinforcement learning (RL) post-training repairs the localized segmental errors of zero-shot text-to-speech (TTS), arriving at a working recipe still relies on tedious manual tuning, and whether LLM agents can take over this research pipeline is unclear. We investigate this question with AgenticTTS-Forge, a collaborative workflow that structures human guidance and agentic execution around a shared workspace, applied to CosyVoice2-0.5B. To measure what the agent automates, we audit its trajectory stage by stage against the published recipe. To measure what it exploits, we score its policies with held-out observers hidden from the agent. Our results show that the agent recovers an underspecified recipe, improves it, and, when gains stall, surveys the literature unprompted and pivots from the LM carrier to the flow carrier, halving Bad cases. However, its autonomy exposes three traps across the data, proxy, and algorithm axes: the held-out set leaks through a channel the contract never reads, a self-shaped reward inflates the proxy where it is scored, and separately tuned policies do not compose additively. These findings show that the binding constraint is measurement rather than reasoning, and can inform the design of harnesses whose contracts read every channel the agent does.

Partial-spoof detectors must reject synthetic content while accepting benign edits. We diagnose temporal anchors and boundary-consistency training using frozen WavLM features, nonlinguistic unit controls, and source-locked evaluation. Genuine-genuine and genuine-fake splices contrast editing false alarms with synthetic-content misses; they do not isolate a unique causal artifact. With layer-6 features, phone units have higher PartialSpoof evaluation frame EER than soft frames (11.18% versus 9.59%). In 500 retrospective PS-eval cases, augmentation reduces genuine-splice false alarms; synthetic-core misses increase at 5% PS-development FPR but not demonstrably at 1%. A 200-case Llama A follow-up retains the false-alarm reduction but does not establish a miss-rate increase. Constructed-case ranking can improve while fixed-threshold misses rise. Consistency gives no uniform gain, and encoder-layer rankings change across corpora. We report frame and event localization with three seeds, including word units, and treat PS evaluation as retrospective. Audit code, selected training scripts, and result summaries are available at https://github.com/mysxs/partial-spoof-diagnostics.

Delayed Streams Modeling (DSM) enables LLM-based streaming automatic speech recognition (ASR) by aligning acoustic and text streams on a common timeline. DSM adds the padding token <p> and the word-start token <w> to the LLM vocabulary and predicts them together with normal text tokens using the same softmax. We first show that <w> can be removed while maintaining competitive recognition performance. Based on this result, we propose Factorized DSM (F-DSM), which separates the waiting probability for <p> from the distribution over the original LLM vocabulary. This factorization removes ASR-specific tokens from the text prediction space and allows the large-vocabulary softmax to be skipped on waiting steps. Experiments on the Corpus of Spontaneous Japanese and LibriSpeech show that F-DSM achieves better recognition performance than DSM. It also greatly reduces GPU memory use while maintaining similar training throughput, provides a small inference speed improvement through softmax skipping, and reduces the degradation in text-only perplexity observed with DSM.

Modern TTS systems increasingly generate synthetic speech at scale for diverse users. This setting calls for content-level provenance that can verify the origin of released speech and attribute it to the requesting user, which generative watermarking can support by embedding multi-bit identifiers directly into synthesized speech. Once released, however, speech may undergo heterogeneous learned transformations during distribution and editing, with reconstruction objectives that can preserve speech utility while affecting watermark recoverability differently. We find that no single watermark carrier remains consistently reliable across reconstruction models, as its survival depends jointly on the embedded structure, reconstruction mechanism, and observation representation. To this end, we propose Thrive, a multi-bit generative speech watermarking framework for modern autoregressive TTS, covering both discrete-token and continuous-representation generation under reconstruction attacks. Specifically, Rise synchronizes watermark injection into intermediate representations with its continued integration into subsequent generation, while Care combines waveform and spectral experts using bit-wise reliability selection. Experiments on both autoregressive paradigms show that Thrive preserves synthesis fidelity, achieves 87.6% average recovery accuracy under reconstruction attacks, and supports source attribution over candidate sets of up to 10,000 identities.

Automatic music annotation is typically tackled under the assumption of a fixed annotation schema. In practice, commercial music catalogs often need to accommodate new musical attributes as needs evolve. Given that expert music annotation is expensive, it is not evident which methodological approach is most effective at accommodating new attributes and backfilling existing tracks; audio-language models promise zero-shot prediction, but at what annotation budget does supervised adaptation become more compelling? We propose a benchmark based on the MGPHot popular music annotation dataset for simulating music schema extension across different annotation budgets. We investigate zero-shot prediction with audio-language models, learning new attributes from pretrained representations, and adapting models trained on existing annotations. Our results suggest that supervised adaptation is more effective than zero-shot prediction even with small annotation budgets, while frozen representation reuse remains the most effective approach for modest budgets without the tuning required by deeper adaptation.

Large Audio Language Models (LALMs) are prone to hallucinating and over-relying on text priors when simultaneously presented with audio and text inputs. To mitigate these hallucinations, we propose utilizing the multimodal Direct Preference Optimization (mDPO) objective, which forces the model to ground its generation in the acoustic input by contrasting intact and distorted audio counterparts. We extend this preference learning framework to the audio domain by applying a variety of acoustic perturbations. Evaluating the Qwen2-Audio backbone across the DCASE 2025 Challenge and AH Existence datasets, we demonstrate that extending mDPO to LALMs significantly enhances temporal reasoning in the complex DCASE dataset, and improves performance on basic existence verification in the AH benchmark. We identify temporal reversal, frequency masking, and random noise as the most effective perturbations. Ultimately, our approach achieves an absolute accuracy improvement of 14.0% on the DCASE 2025 dataset and 27.4% on AH Existence.

Multi-pitch estimation in vocal ensembles is challenging because singers occupy overlapping pitch ranges and often sing at closely spaced fundamental frequencies, causing their harmonics to overlap in time-frequency representations. Existing models commonly use harmonic constant-Q transform (HCQT)-based representations to provide frequency-adaptive resolution, at the cost of expensive feature extraction when training mixtures are generated on the fly. We revisit this design and compare HCQT with a linear short-time Fourier transform (STFT), whose frequency bins are directly provided as model inputs. Despite its fixed frequency resolution and the absence of a pitch-aligned input grid, the linear STFT outperforms HCQT while substantially reducing feature-extraction cost. Further analysis shows that a longer analysis window or broader spectral coverage provides no additional improvement, while restricting the input to the predicted pitch range reduces the advantage of the linear STFT. These results suggest that finer frequency resolution does not necessarily improve vocal-ensemble MPE, and that shorter analysis windows can be more effective for time-varying vocal pitches.

Spatial semantic segmentation of sound scenes (S5) involves the detection and extraction of target sound events in audio files. However, in typical detection-extraction pipelines, errors in the detection module propagate to the extraction module, degrading overall performance. In this work, we develop a multichannel detection-extraction model and evaluate inference-time algorithms to improve the class detection and target sound extraction in a multi-stage setup. We evaluate two fitness functions: binary cross entropy and mixture consistency; two relabeling strategies: naive and conditional confusion matrix relabeling; and three source estimate evaluation models: a multi-label classifier, a fine-tuned single-source classifier, and an off-the-shelf audio judge. We thoroughly evaluate the impact of these design choices on the DCASE 2025 Task 4 dataset, and validate entropy-based gating of inference-time updates on the held-out evaluation set, laying the groundwork for inference-time updates of S5 systems in real-world scenarios.

Post-training text-to-music generation requires reward signals that capture multiple aspects of musical quality beyond what any single automatic metric can measure. We study structured, rubric-based rewards from pretrained audio-language models (ALMs) as training signals for both autoregressive and diffusion-based music generators. An ALM scores each generated clip against the rubric; we rank candidates generated for the same text prompt by their scores and convert these rankings into preference pairs for DPO on both MusicGen-small and ACE-Step v1, and additionally use the rubric scores directly as scalar rewards for DiffusionNFT on ACE-Step v1. On MusicCaps, rubric-based optimization improves CLAP, SongEval, and Audiobox-Aesthetics simultaneously, with the strongest gains obtained by DiffusionNFT on ACE-Step. By contrast, on MusicGen-small, building preferences from any one of these automatic evaluators produces clear cross-metric trade-offs: the targeted evaluator improves while other independent evaluators deteriorate. We further study tempo, key, and instrumentation, where precise objective rewards are available. Directly optimizing these specialized rewards reliably improves the target attributes, whereas ALM rubrics provide only partial transfer for tempo and instrumentation and no measurable improvement for key. Together, these results suggest a practical division of labor: ALM rubrics are effective for broad perceptual qualities that are difficult to formalize, while specialized objective rewards remain preferable when reliable measurements are available.

Music information retrieval often relates signal-derived, algorithmic, and symbolic information across different coordinate systems. Existing visualizations typically leave notation separate from physical time, while composites that combine them are assembled by hand. We present LayerIt, an open-source Python library for composing independent representations with notation on a shared performance-time axis. Given a score warped into performance time and a note-level alignment, LayerIt emits a single SVG that preserves the score's MEI structure and keeps added components identifiable and restylable. We demonstrate this through a challenging beat-tracking analysis, in which tracker output, signal representations, and notation can be inspected together to locate metrical disagreement and other errors against both notated structure and performed time.

This paper introduces PEACE, the first joint embedding of audio effect code and output audio. Building on SLAP's multimodal objective, we pair an AFx-Rep audio encoder with two code encoders for Faust, a functional language for audio signal processing. First, we evaluate a fine-tuned T5 transformer over Faust source code. Second, we evaluate a message-passing graph neural network over an intermediate representation of the Faust compiler, capturing both topology and UI parameters. We evaluate on audio-to-code retrieval, where masking UI parameters at inference yields embeddings that encode effect chain topology alone. When parameters are visible, the two code encoders tie on retrieval of mixed-length chains but have tradeoffs on single-effect galleries. With parameters fully masked, PEACE recovers ordered chain topology far above chance without the limitations of supervised methods. PEACE outperforms pretrained models on an out-of-distribution reverb retrieval benchmark and can improve frozen audio-only representations. Its dual understanding of topology and parameters lays the groundwork for music information retrieval systems that search, generate, and condition on DSP code.

This paper presents a web-based application for navigating and composing microtonal harmony, built on the Harmonic Eigenspace, a four-dimensional psychoacoustically grounded space in which tetrad chord types are located by their spectral dissonance profiles, computed with Sethares's roughness/dissonance model. The coordinate system is transposition-invariant: a coordinate triple ({\alpha}, e̱ta}, {\gamma}) locates the three upper notes in relation to the root, so a chord quality corresponds to a direction in the space, the invariant ray along which transposition acts, while the root frequency sets the scale. The dissonance field over these coordinates can be computed at any register; the locations of its local minima are register-invariant, as they arise from partial-coincidence ratio conditions. The dissonance volume contains 100 local minima that align with just-intonation intervals and act as landmarks, organising the space into basins around the most consonant tetrads. We embed tetrads from three tonal equal temperaments as discrete lattices within this continuous volume. The application presents this space through two components: the Harmonic Eigenspace as a navigable 3D visualisation of the dissonance volume in which all nodes are playable, and a Modal Studio that extends modal interchange logic to the ten-gradation interval vocabulary of 53-TET. The application also functions as a MIDI controller with MIDI Polyphonic Expression support, usable in any digital audio workstation that supports this format. A listening study with 31 participants used both scenes of the application: listeners first rated isolated 53-TET chords alongside chords familiar from Western practice, such as the maj7 and the m7; they then rated chord progressions composed in the Modal Studio, measuring their acceptance or rejection of microtonal progressions heard for the first time.

Few-step neural text-to-speech models often rely on short- ened diffusion or flow-matching schedules, or on distillation from pretrained multi-step teachers. To avoid these depen- dencies, we present DriftTTS, a few-step mel-spectrogram generator trained without a generative teacher, distillation, or adversarial discrimination. DriftTTS uses a distribution- matching drift objective in a mel-domain feature space defined by raw mels and a frozen masked-autoencoder encoder pretrained on the same LJSpeech training split. On-policy rollout trains the decoder on its own interme- diate states and supports inference up to the trained roll- out depth. On LJSpeech, DriftTTS at NFE=4 achieves 3.87 dB MCD and 3.7% WER, compared with 3.85 dB and 3.4% for Matcha-TTS. In a fully paired blind listen- ing test, DriftTTS obtains 4.18 MOS, compared with 3.96 for Matcha-TTS and 4.22 for ground truth. These results demonstrate competitive few-step synthesis without a pre- trained generative teacher. Code can be found at https: //github.com/BASHLab/driftTTS.git

Speech delivery varies with both the requested paralinguistic attribute and the linguistic content. We introduce ParaGeo, a matched-content decomposition of paralinguistic variation in a frozen speech language model. Synthesized audio tokens are replayed with a fixed listening prompt; pooled key/value (K/V) representations are centered and projected into a shared low-dimensional space. Our GLM-4-Voice probe spans 80 requested controls from 12 benchmark families across eight sentences. With a globally fitted calibration basis, content-held-out centroid accuracy using this basis is 9.49% versus a 1.25% permutation baseline; same-label cross-content cosine similarity is 0.285 versus 0.017, and both conditional permutation tests yield p = 0.001. A separate ten-scenario, six-style probe reveals reproducible contrast directions across scenarios. Static, additive, and temporal interventions produce attribute-, layer-, and schedule-dependent response profiles. These results provide a shared coordinate representation for measuring paralinguistic structure and an empirical starting point for latent speech control. Code is available at https://github.com/yuhanlydia/ParaGeo.

Knowledge-driven neural vocoders struggle to learn reliable fundamental frequency end-to-end, because spectral objectives provide weak supervision of periodic structure and lack phase information. We address this with a source-filter model whose alias-free additive source makes the instantaneous phase of the glottal cycle explicit; differentiating it yields the instantaneous frequency, and thus $F_0$, without an external tracker. Waveform error supervises only the deterministic harmonic path, while a spectral loss covers the full signal. On M4Singer and LM-SSD, the reconstruction is phase-aligned, reaching a signal-to-reconstruction-error ratio of 8.1 dB, while neural baselines remain negative. However, GOLF, given an external $F_0$, still reaches lower spectral distortion. On LM-SSD, the recovered $F_0$ attains the highest overall accuracy of any method tested, including supervised neural pitch trackers applied off the shelf, and the glottal closure instants come within 0.53 points of REAPER's identification rate, without any $F_0$ label.

This work presents an automatic speech recognition (ASR) system personalized for a Czech speaker with a permanent tracheal stoma and severe dysarthria rendering their speech unintelligible to untrained listeners. We release a public dataset containing 33 annotated hours of the speaker's speech, collected using a novel "artificial conversation" protocol designed for high engagement and dialogue realism. We propose a multi-stage training pipeline based on Whisper Base: fine-tuning on standard Czech speech, acoustically simulated tracheostomic speech, and the speaker's data. We evaluate the system across three near real-time scenarios: scripted conversations, question answering, and spontaneous dialogue, achieving a 50% relative reduction in Character Error Rate compared to Whisper Base baseline and surpassing the average recognition accuracy of their assistants in acoustic recognition of isolated utterances. We demonstrate that even for severely impeded speech, a helpful ASR is achievable, as evidenced by the quantitative results and the feedback from the speaker.

Real-time conversational agents require speaker diarization that streams and runs on a CPU. Most systems apply an utterance-level speaker encoder to short, heavily overlapping chunks, which wastes computation and leaves the model optimized for the wrong task. We instead turn a state-of-the-art speaker recognition architecture into a causal frame-level encoder that reads the stream once and emits one embedding every 80~ms from a bounded two-second window of past audio, and pair it with online clustering that gates every update on the self-similarity of the stream. Trained only by distillation from an utterance-level teacher on simulated and out-of-domain mixtures, and evaluated with one fixed set of hyperparameters, the system is the most accurate streaming diarizer on low-overlap benchmarks at sub-second latency, degrades far less than cache-based systems as the number of speakers grows, and runs five times faster than real time on a single CPU thread.

Jazz pianists develop distinctive traits that experienced listeners can often identify within seconds, yet the features underlying this recognition resist formal description. We study jazz pianist style through the lens of a pretrained symbolic music transformer, showing that its learned representations already encode pianist identity well enough for highly accurate classification across two benchmarks. We then augment the transformer with cross-attention over learned pianist identity embeddings, enabling it to generate music conditioned on a specific artist's style. Two evaluation protocols confirm that the generator captures meaningful stylistic structure: a sliding-window classifier consistently attributes conditioned continuations to the correct artist, far above unconditioned baselines; and a classifier trained entirely on synthetic generations identifies real pianists across 12 classes with 87% chunk-level and 95% song-level accuracy. Finally, we repurpose the classifier to locate the most characteristic moments within a performance, surfacing the specific musical gestures that distinguish each pianist's voice.

We investigate how predictive performance depends on the maximum context available to an autoregressive model of symbolic music, and what information long-context models exploit. A small causal Transformer is trained separately at each context length T in {6, 18, 48, 96, 192, 336} and evaluated on the same target positions, both over raw tokens and over non-overlapping binary latent codes. On Nottingham folk tunes, the token predictor's test NLL decreases by 71% (0.986 bits per token) between 6 and 336 tokens, and by 71% on the O'Neill's tunes long enough for the same sweep (129 of 302 test tunes). Most of the long-context gain is explained by exact copying: the improvement appears when an earlier occurrence of the target's 16-token history enters the available context; overwriting that occurrence removes the gain, whereas equally large unrelated corruption does not; and a simple copy baseline recovers 94% of the reduction. A trained 336-token model likewise loses most of this gain when its history is restricted to recent tokens at test time. Because many repeats arise from written repeat signs expanded in the score, this result pertains specifically to these rendered score representations. By contrast, on MAESTRO performances and MusicNet scores, where exact repeats are substantially less frequent, the reduction is smaller (9% and 18%) and is largely attained by 96 tokens. Finally, an audit of an earlier draft that reported saturation at 16 tokens identified split leakage, overlapping latent receptive fields, an averaging predictor, and a per-file tempo grid; we present these as methodological checks for context-length measurements.

A single compact encoder for both speech and music removes the need to maintain a separate model per domain. A practical recipe distils a speech teacher and a music teacher into two small students, then merges them into one model. Two merging algorithms exist: interpolating the student weights from a shared initialisation, or permuting their channels into alignment before averaging. Prior work fixed the distillation loss in both, leaving open whether the loss itself affects how well the students merge. We show that it does. Holding architecture and evaluation fixed, we replace the DistilHuBERT loss \Lkd{} with a correlation-based loss \Lcl{}. Under weight interpolation, \Lcl{} students stay closer to their own better endpoint in $17$ of $18$ comparisons and lead on the speech tasks at every interpolated weight. Under activation permutation, they need fewer channels rearranged at three layers, their matched channels correlate more strongly, and they again lead on speech. The two merging algorithms share no mechanism, yet both improve under the same loss substitution. This suggests that the gain lies in the representation \Lcl{} produces rather than in either algorithm.

Self-supervised learning (SSL) speech models are accurate but large. Knowledge distillation compresses them, but the student loses the teacher's noise robustness. Correlation-based distillation addresses this with two terms: cross-correlation aligning student and teacher's representations, and self-correlation decorrelating the student's features. Both the original method and De'HuBERT credited the self-correlation term without isolating it. We show the opposite. On LibriSpeech-100 distillation with held-out CHiME-3 noise at 10\,dB, an unbiased full-dimensional probe, a Pearson-variance decomposition, a same-noise causal control, and a per-dimension analysis identify the cross-correlation diagonal as the mechanism that encourages noise invariance, lowering noise-classification accuracy from 76.98% to 55.02%, whereas adding the self-correlation term back leaves it at 55.55%. The self-correlation term instead reorganises the feature space and improves accuracy on the clean downstream tasks, but removes essentially no noise. Across nine speech and music tasks this yields a concrete design rule for distillation: weight the cross-correlation term for noise robustness, tune the self-correlation term for better downstream performance, and sample teacher and student noise independently.

Multi-scale design is crucial for efficient audio-visual speech separation, yet effectively modeling multi-scale information for audio-visual feature fusion remains challenging. We argue that the limited capacity of existing approaches primarily arises from: 1) treating features from different modalities in the same manner, and 2) overlooking the role of global features. To address these issues, we propose a Global-guided Asymmetric Attention Network (GAANet). Our model introduces two core innovations: first, an asymmetric multi-scale fusion framework that allows audio and visual streams to extract and interact with features at their respective optimal temporal resolutions, removing the need for symmetric temporal downsampling; second, a global-guided attention mechanism that compresses each modality into a compact global token with a temporal dimension of one, which then provides high-level semantic cues to guide both intra- and inter-modal fusion across scales. Experiments on LRS2 and VoxCeleb2 demonstrate that GAANet achieves state-of-the-art performance, reaching 16.5 dB SI-SNRi on LRS2 and 14.0 dB on VoxCeleb2, while maintaining a lightweight computational profile with only 3.3M parameters and 19.8G MACs. These results highlight the strong potential of asymmetric temporal modeling and global guidance for efficient and robust multimodal fusion. The source code is publicly accessible at https://github.com/redizzy/GAANet

Spoofing-aware speaker verification (SASV) must confirm who is speaking and that the speech is genuine, but current systems gain one at the expense of the other. Modular cascades are accurate yet run two large self-supervised encoders, whereas end-to-end models are efficient but lose accuracy when a single embedding has to serve two conflicting objectives. We show that this trade-off between efficiency and specialization is avoidable. A single frozen W2V-BERT~2.0 backbone is adapted per task by deep hybrid Wavelet Prompt Tuning (WPT), so each branch obtains its own view of the shared encoder through dedicated prompts and a task head, and the scores are combined only at inference. Training only about 7M parameters, a small fraction of those a strong two-encoder cascade requires, the system surpasses it on SpoofCeleb evaluation with 0.03% CM-EER and 0.038 min a-DCF against 0.16% and 0.047. Because the backbone is shared and frozen, new branches attach without retraining the heads and prompts already in place. On ASVspoof5, with adversarial attacks and codec distortions, it reaches 0.092 min a-DCF and 3.60% CM-EER using only the provided data, indicating that the design holds under realistic in-the-wild threats.

Detecting synthetic and voice-converted speech remains difficult for low-resource languages with dialectal diversity, where systems must generalize across regional dialects and unseen acoustic channels. We present the UZH-CL submission to the ArA-DF 2026 Shared Task on Arabic speech deepfake detection, covering Track~1 (dialect generalization) and Track~2 (acoustic robustness). We freeze a W2V-BERT-2.0 backbone and adapt it with Wavelet Prompt Tuning, updating under 1% of parameters, and aggregate multi-layer representations with cross-layer attention and a general attentive-statistics pooling head rather than a specialized graph backend. Complementary detectors are obtained by varying adaptation strategy, training data, augmentation, and encoder family. We find that the two shift types require different fusion regimes: a broad multi-window ensemble for dialect generalization, and a compact, channel-matched, center-crop ensemble for acoustic robustness. Official evaluation yields 1.96% EER on Track~1 (6th place) and 1.04% EER on Track~2 (3rd place), corresponding to 87% and 96% relative reductions over the XLS-R+AASIST baselines.

Increasing depth of transformer models improves recognition, but it comes at a substantial cost. Each additional layer requires more parameters, which makes the process computationally inefficient. We ask whether additional processing can focus on integrating features already computed. Looped Audio Spectrogram Transformer (LAST) first processes all tokens, then reuses the same blocks to refine only the class token over fixed audio features, thereby making later passes inexpensive. On AudioSet, ten-pass LAST achieves 0.345 mean average precision, exceeding a twelve-layer sequential transformer by 2.1% relative with 49.4% fewer parameters, 42% fewer multiply-accumulate operations, and 9.8% higher measured throughput. Across separately trained models, increasing the pass count from two to ten improves accuracy while adding only 1.2% computation. Further evaluations show improved robustness to temporal masking and various other auditory augmentations, with better generalization on classification tasks with music, environmental, and event sounds.