Daily Papers — 2026-08-25

21 papers on audio, speech, music, and acoustics.

1. TurnBench: A Multi-Domain Benchmark for Turn-Taking Dynamics in Spoken Dialogue

Authors: Freeman Jiang, Ramon Sanabria, Soham Deshmukh, Bandhav Veluri, Simon Michael Vuch Williams et al.

Categories: eess.AS, cs.CL | 8 pages, 2 figures. Submitted to IEEE SLT 2026 Score: 7.79/10 (Obj:8 Id:8 Ind:8 Comp:8 Eff:9 Nov:7 Rep:6)

Strength: 30 小时三重标注六类型双语料 + 104h 训练集 + 公开排行榜，统一协议评测 14 个系统，给出 EOT 跨类型稳定、INT 假阳性集中于 backchannel 密集类型、人类提前 151ms 转接而最强系统 VAP 也需 994ms 延迟的量化结论。
Weakness: 论文承诺的 GitHub 评分仓库（SesameAILabs/turnbench）当前返回 404，核心 scorer 不可获取；全双工模型评测依赖输出音频反推决策且只覆盖 EOT。
Full review: Claude Code 全文七公理审稿

Speakers in natural conversation take turns speaking and listening, deciding in real time when to take, hold, or yield the floor. However, turn-taking evaluation remains limited due to the lack of a consistent, linguistically grounded evaluation protocol and hand-annotated data covering diverse conversation types. To address this, we present TurnBench, a multi-domain benchmark that pairs a 30-hour, hand-labeled corpus of dyadic human conversation with a standardized evaluation protocol for end-of-turn and interruption detection. We set conversation type as a controllable experimental variable, covering six distinct interaction styles, and triple-annotate each conversation. Benchmarking 14 heterogeneous turn-taking systems, we find end-of-turn recall stable across types, while interruption false positives are strongly type-dependent and concentrated in backchannel-dense interaction styles. Although in smooth floor transfers human listeners begin speaking a median 151 ms before the current turn ends, no current system performs equivalently without incurring excessive false positives. We release our corpus, a 104-hour training set, and a public leaderboard with an interactive dataset viewer at https://turnbench.sesame.com

2. FireRedAudio: A General-Purpose Audio Language Model with Decoupled Continuous Representations for Understanding and Generation

Authors: Junjie Li, Xuelong Geng, Kun Xie, Feiyu Shen, Yichen Wu et al.

Categories: cs.CL, cs.SD | 20 pages, 3 figures Score: 7.60/10 (Obj:8 Id:8 Ind:6 Comp:6 Eff:8 Nov:7 Rep:8)

Strength: 共享 9B LLM 上以解耦连续表征统一六类音频任务，MMAU test 80.9 / MMSU 83.3 / FLEURS-102 平均 14.94% / Seed-TTS-Eval 平均内容错误 1.20% 领先，Instruct TTS 六项全胜，权重与代码开源。
Weakness: 全文无消融验证双路径设计相对单路径/融合方案的收益，长音频时间接地的 50 条评测由对比对象 Qwen3.5-Omni-Plus 自裁、评测集与训练构造同构，且缺少推理成本与人类感知评测。
Full review: Claude Code 全文七公理审稿

A unified audio model must recognize and understand linguistic, paralinguistic, and environmental information while supporting speech synthesis and editing. A key challenge is representation: understanding favors compact features suited to long-context modeling, whereas speech generation requires reconstructible features that preserve fine-grained acoustic detail. We introduce FireRedAudio, a general-purpose audio language model with a shared 9B-parameter LLM. To the best of our knowledge, it is the first publicly disclosed unified audio-language model to provide separate continuous input representations for understanding and generation within a single trainable autoregressive LLM. Audio to be recognized or analyzed is processed by a dedicated Audio Encoder, while speech inputs for generation use a RedAE-based pathway. The LLM directly generates text or conditions a flow-matching DiT to produce continuous acoustic latents. Through progressive multitask training, FireRedAudio supports ASR and audio understanding, with the latter extending to recordings of up to one hour, as well as zero-shot TTS, Instruct TTS, and semantic and acoustic speech editing. Its structured organization of long-form audio achieves second-level timestamp accuracy. Across comprehensive evaluations, FireRedAudio achieves competitive or leading performance in audio understanding and multilingual ASR, strong content accuracy and speaker preservation in zero-shot TTS, leading instruction following in Instruct TTS, and substantial improvements over Ming-UniAudio-Edit in both semantic and acoustic speech editing. These results demonstrate the viability of decoupled continuous input representations for unifying audio understanding and continuous-latent speech generation in a model of moderate scale. Our code is available at https://github.com/FireRedTeam/FireRedAudio.

3. Can We Read the Mind of an Audio LLM? A Verbalizable, Multilingual Middle-Layer Workspace

Authors: Jiajun Fan, Jingyuan Li, Prashanth Gurunath Shivakumar, Qi Luo, Jia-Hong Huang et al.

Categories: cs.SD, cs.AI, cs.CL Score: 7.37/10 (Obj:9 Id:8 Ind:8 Comp:9 Eff:7 Nov:7 Rep:4)

Strength: 用波形交换控制把声波驱动信号与文本先验分离，在 48 层 Qwen3-Omni 上读出隐藏多跳链（转录为乱码的 Nixon clip 重建 water+gate→scandal→president→Nixon），工作空间带真实 vs 静音 40.0% vs 21.8%（p=4.7×10⁻⁵），38.5% top-1 读出为中文并附频率对照，幻觉中位 12% 深度早生。
Weakness: 无任何代码或 curated 数据发布（curation 清单与人工概念标注不可复现），监督效用止于现象展示、无端到端拦截演示，40% 平衡准确率与 4-30% 信号稀疏度离部署级监控有实质距离。
Full review: Claude Code 全文七公理审稿

An audio language model is a black box in a specific way: we see what it says, never what it works out on the way there, and chain-of-thought monitoring helps only if the model writes its reasoning down. Reading a base Qwen3-Omni with a logit lens at the audio-token positions, we find that the answer to a spoken question becomes legible - in words - in the model’s middle layers, before it emits any token. Five findings follow. (1) The readout carries concepts in neither the question, the options, nor the model’s own transcription: on a clip whose verbatim transcription is empty garbling, it reconstructs Watergate and scandal, passes through the role president, and resolves to Nixon - a hidden multi-hop chain, read with no chain-of-thought. (2) The content is language-agnostic: one audio-inferred concept surfaces in several scripts at once, and 38% of top-1 readouts are Chinese on English inputs. (3) It is paralinguistic: given the same clip as audio and as the model’s own emotion-free caption, the audio mind forms the sound source, speaker role, or affect that the caption discards, and answers correctly more often. (4) The audio-driven signal is absent at the input, turns on about a tenth of the way into the network, separates most cleanly from the text prior in the middle band (35-80% of depth), and activation patching shows it is causally used and committed before the last fifth of the layers. (5) Deleting single layers maps the pipeline: reading the sound in is localized to the entry layers and answer delivery to the output layer, while retrieval is distributed across the interior. Throughout, a waveform-swap control - identical text, only the sound changed - isolates the audio-driven signal from a prior over the printed options. This is a qualitative account of what an audio model works out before it speaks: the quantities are controls, not benchmark scores.

4. LibriBrain100: One Hundred Hours of Broad and Deep MEG Data for Neural Speech Decoding at Scale

Authors: Francesco Mantegna, Dulhan Jayalath, Gereon Elvers, Tasha Kim, Benjamin Ballyk et al.

Categories: cs.LG, cs.CL | 10 content pages + 22 pages for references and appendices (in total: 32 pages, 16 figures, 6 tables) Score: 7.32/10 (Obj:9 Id:7 Ind:7 Comp:6 Eff:8 Nov:6 Rep:9)

Strength: LibriBrain100 提供 104 小时 MEG（单受试者约 80 小时，为次深数据集的 8 倍）并配套标准划分与开放工具；把 sub-0 的 80 小时纳入联合训练使 32 名受试者的跨用户词分类泛化提升约 15 个百分点，微调数据降至 25%（约 10 分钟）仍无显著退化。
Weakness: 数据质量验证依赖本组自研 MEG-XL 骨干且止步于 50 词 top-10 平衡准确率，无脑-文本基线；32 名受试者共享同一两章刺激，跨受试者泛化仅在极窄内容域成立，数据量消融全部不显著。
Full review: Claude Code 全文七公理审稿

We introduce LibriBrain100, a large-scale MEG dataset for speech decoding designed from the ground up for reproducible, standardised evaluation. LibriBrain100 more than doubles the size of the original LibriBrain release, resulting in over 100 hours of high-quality MEG acquired while subjects listened to naturalistic continuous speech. With $\sim$80 hours from a single subject, LibriBrain100 sets a new record for deep, within-subject neural data (8$\times$ more than the next comparable dataset and roughly 80$\times$ more than other datasets). To demonstrate the payoff of this depth-first design, we evaluate on a word-classification benchmark—an increasingly well-established stepping stone towards the open challenge of noninvasive brain-to-text decoding. Using an existing decoding model, we achieve state-of-the-art performance—validating both the quality of the recordings and the value of within-subject data at scale. Because collecting 80 hours of data per user is impractical for real-world applications, we also collected $\sim$40 minutes of additional data from each of 32 subjects. Using the same word-classification benchmark, we demonstrate the value of broad multi-subject data: supervised finetuning of a pre-trained model can substantially compensate for limited per-subject data. We provide standard train, validation, and test splits, all reproducible through an open-sourced Python library that supports easy downloading, optional preprocessing, and data loading for common deep learning frameworks. In addition, the dataset and evaluation infrastructure are being released alongside an open machine-learning competition with a public leaderboard for standardised benchmarking. Ultimately, our hope is that LibriBrain100 will accelerate progress towards practical non-invasive brain-computer interfaces, capable of restoring communication to people living with severe paralysis.

5. SPECTRA: Subspace-Preserving Embedding Calibration, Transport, and Replay for Fully Few-Shot Class-Incremental Audio Classification

Authors: Giries Abu Ayoub, Loay Mualem, Simon Korman

Categories: cs.SD, cs.AI Score: 7.30/10 (Obj:8 Id:8 Ind:7 Comp:6 Eff:8 Nov:8 Rep:4)

Strength: 冻结 PENGI 上叠加 adapter + 子空间特征回放 + 转导 OT，三基准 50 seeds 平均准确率全面超越 TAPE（LS-100 AA +7.09、NSynth-100 +3.04、FSC-89 +1.88），遗忘全面下降，10-way 硬协议同样全胜，Gaussian/子空间消融证明增益来自子空间结构而非回放行为。
Weakness: 全文无代码/数据/权重发布（GitHub 检索为空），转导 OT 用测试批自精炼原型且是唯一有 PD 代价的组件（+0.3/+0.5 AA 对 +0.4/+0.9 PD），净收益与跨批稳定性缺证据。
Full review: Claude Code 全文七公理审稿

Fully few-shot class-incremental audio classification (FFCAC) requires recognizing new sound classes from only a handful of labeled examples per session, without forgetting previously learned classes and without any large base dataset. Existing methods typically freeze a pre-trained audio–language encoder and classify with point prototypes, but they suffer from significant performance degradation throughout the sessions due to generic feature representations. We propose SPECTRA, a framework built on a frozen encoder which adds three components. (i) a lightweight trainable adapter that calibrates the generic embeddings to the task; (ii) subspace feature replay, an exemplar-free anti-forgetting scheme that replays old classes by sampling from the low-rank subspace of their stored features; and (iii) a transductive optimal-transport refinement of prototypes at test time. Our central finding is that the subspace structure of the replay diminishes forgetting and outperforms naive Gaussian replay of equal variance. On three FFCAC benchmarks (NSynth-100, FSC-89, LS-100), SPECTRA improves average accuracy and reduces forgetting over current state-of-the-art methods, and our ablations statistically validate each component.

6. TurboT2VA: Fast Large-Scale Text-to-Video-Audio Generation via Score-Regularized Consistency Distillation

Authors: Xiaoda Yang, Yuxiang Liu, Kaiwen Zheng, Yuan Liu, Yibo Lai et al.

Categories: cs.CV Score: 6.95/10 (Obj:1 Id:2 Ind:1 Comp:1 Eff:2 Nov:2 Rep:1)

Strength: 19B 联合 T2VA 模型四步蒸馏 + 模态感知推理栈实现 20.1×（512×768）与 54.67× generator-only（1024×1792, 单 H20）加速，逐组件延迟拆解与 13 基线对比完备，推理代码与权重已开源。
Weakness: 方法论是 rCM/DMD/sCM 的有序编排而非新目标函数，缺等价成本对照下的独立效用边界，训练脚本未发布，且音频指标存在 4.902→4.486 的可测回退。
Full review: Claude Code 全文七公理审稿

Joint text-to-video-audio generation produces synchronized visual and acoustic content, but the long sampling trajectories and heterogeneous multimodal computation of large models make inference prohibitively expensive. We present TurboT2VA, a distillation and inference framework for accelerating a 19B-parameter joint video-audio model. Large-scale T2VA distillation is challenged by modality-imbalanced optimization, the difficulty of continuous-time consistency training at scale, and the quality–diversity trade-off. TurboT2VA addresses these issues with per-modality normalization and a progressive curriculum comprising discrete consistency warm-up, continuous consistency refinement, and joint consistency–distribution matching. The curriculum first establishes a stable, diverse generation trajectory and only then introduces distribution-level refinement. On LTX-2, four-step distillation reduces generator latency from 50.52s to 2.51s at the standard evaluation resolution of 512$\times$768, achieving a 20.1$\times$ speedup while maintaining strong visual quality, audio fidelity, diversity, and video-audio synchronization. We further develop an architecture-aware inference stack that combines guarded W8A8 and fused operators, padded-text compaction, and modality-aware sparse attention while preserving dense cross-modal and text-conditioning paths. Under the high-resolution deployment setting at 1024$\times$1792, the complete stack reduces generator latency from 318.74s to 5.83s on one NVIDIA H20, achieving a 54.67$\times$ generator-only speedup. Inference code and generation demos are available at https://github.com/thu-ml/TurboDiffusion/tree/main/turbot2va.

7. Lost in Speech: Trilingual Spoken Hallucination Detection Across Audio and Transcripts

Authors: Meruyert Aristombayeva, Jason S. Lucas, Chaewan Chun, Dongwon Lee

Categories: cs.CL Score: 6.90/10 (Obj:1 Id:2 Ind:1 Comp:1 Eff:2 Nov:2 Rep:1)

Strength: 首个英/俄/哈三语口语幻觉基准，含 12,013 条合成样本与 290 条 factcheck.kz 真实谣言评估集；合成训练检测器向人类书写虚假内容迁移良好（真实 macro-F1 0.82-0.88），并以 2×2 溯源矩阵定量拆解了合成基准的 provenance 混淆。
Weakness: 数据集”录用后公开”、无代码仓库，当前不可复现；零样本音频检测绝对性能接近随机，且哈萨克语 TTS-ASR 管道 38.3% 重大语义偏差使低资源信号混杂级联噪声，削弱结论效度。
Full review: Claude Code 全文七公理审稿

While text-based hallucination detection has been extensively studied, spoken hallucination detection remains largely unexplored, particularly for low-resource languages. We present the first multilingual spoken hallucination benchmark comprising 12,013 news samples across English, Russian, and Kazakh with controlled hallucinations of three types and three severity levels. Samples comprise original articles and aligned hallucinated counterparts in text and audio. We complement the synthetic corpus with 290 fact-checked fake news items collected natively in Russian (225) and Kazakh (65), translated into the other language and rendered through the same TTS-ASR pipeline. We assess fine-tuned multilingual encoders and, in zero-shot in-context settings, multimodal decoder models on transcript-based versus direct audio processing. Transcript-based detection generally outperforms direct audio processing, with binary-task degradation for strong encoders tracking per-language ASR error. On real-world fakes, synthetic-trained detectors transfer strongly (macro-F1 0.82-0.88 on original text), while Russian provenance analysis reveals both veracity-related and model-dependent machine-style signals, quantifying a key confound in synthetic hallucination benchmarks.

8. REDnet: Recursive Encoder and Decoder for Speech Separation under Unknown Number of Speakers and Variable Number of Microphones

Authors: Fulin Wu, Zhong-Qiu Wang

Categories: eess.AS | in submission Score: 6.53/10 (Obj:1 Id:2 Ind:1 Comp:1 Eff:2 Nov:2 Rep:1)

Strength: 首个用递归编码器+递归解码器在单一 DNN 中同时处理未知说话人数与可变麦克风数的分离系统，与 FlexIO 同协议对比全面占优（WHAMR!-R 双麦 15.69 vs 12.5 dB SDR），5-mix 最难点 20.8 dB 为新 SOTA。
Weakness: 全文无代码/权重/数据发布，主要对比锚点 FlexIO 尚未公开、本文当前 0 引用且无第三方复测，真实录音泛化仅凭 DNSMOS 单一无参考指标。
Full review: Claude Code 全文七公理审稿

We propose $\textit{recursive encoder and decoder}$ (RED) for building a single deep neural network (DNN) model that can separate multi-speaker mixtures containing unknown numbers of speakers and variable numbers of microphones arranged in an unknown geometry, a task that has not been studied yet. The decoder of RED recursively detects whether there are active speakers left and separates one speaker at a time. It is designed to be trained in an end-to-end fashion to improve separation performance. The encoder of RED recursively encodes each microphone channel of the input mixture, sequentially incorporating spatial cues. Combining both, the DNN can be trained to separate mixtures not only with unknown numbers of speakers but also with variable numbers of microphones, achieving state-of-the-art performance on multiple public datasets.

9. SeMoCo: A Semantic-First Motion Codec for Motion Language Modeling

Authors: Tianlv Huang, Hetian Guo, Ziyi Cai, Song Wang, Yanping Zhang et al.

Categories: cs.CV, cs.CL, cs.GR Score: 6.32/10 (Obj:8 Id:6 Ind:8 Comp:5 Eff:5 Nov:6 Rep:8)

Strength: SeMoCo 在共享 split 重建评测中以 12.83mm MPJPE-22 大幅领先 MoMask/MotionMillion 的 78.95–96.01mm，且语义监督置于并行分支把对齐代价从 2.65mm 压到 0.40mm（表 11）。
Weakness: 语义优先的核心收益只在自建 TMR-SOMA track 上成立，其生成侧 FootSkate .081 vs Kimodo .028、Jerk 874 vs 81，标准 HML-263 协议下 FID 8.75 落后主流方法，跨 source 增益不稳定。
Full review: Claude Code 全文七公理审稿

Discrete motion representations have substantially advanced autoregressive text-to-motion generation. However, most motion tokenizers are optimized for reconstruction and do not explicitly allocate capacity according to semantic role. Action-level meaning and fine-grained kinematic detail must therefore be encoded through the same reconstruction-driven hierarchy. We introduce SeMoCo, a semantic-first motion codec, together with a dual-axis motion generator for language-conditioned motion generation. Each motion token contains one semantic token and a residual sequence of kinematic tokens. The generator models semantic progression across time and autoregressively refines the residual entries. We also construct $Ω$-MotionVerse, a large-scale, multi-source human-motion dataset unified under the SOMA representation. Across the reported comparisons, SeMoCo achieves the best reconstruction accuracy among the compared codecs, while strong text-to-motion results demonstrate the effectiveness of its motion tokens for downstream generation.

10. ROMNet: a hybrid reduced order modeling and machine learning approach to waveform inversion

Authors: Liliana Borcea, Alexander Mamonov, Kui Ren, Haizhao Yang, Chugang Yi

Categories: math.NA, cs.LG, physics.geo-ph Score: 6.21/10 (Obj:8 Id:6 Ind:6 Comp:8 Eff:6 Nov:6 Rep:4)

Strength: ROMNet 将 ROM 波形反演的单次 Gauss-Newton 迭代成本从 19.4s 降至 0.26s（端到端 14min 降至 10.5s，约 80 倍加速），在 Random Gaussians（rel. L2 0.0297）与 GeoFWI（0.0358）上均优于 Fourier-DeepONet 与 InversionNet。
Weakness: 精度未超过其所加速的直接 ROM 反演（两类数据 ROM 均相当或更低），且 GeoFWI 对比中 ROM 反演获得了平滑真实速度作为初始猜测，代码与权重未发布。
Full review: Claude Code 全文七公理审稿

Waveform inversion seeks to estimate the wave speed of a heterogeneous, inaccessible medium, from time-resolved measurements of the waves at user controlled sensors. We consider this inverse problem for acoustic waves and an active array of source/receiver sensors that emit probing signals and measure the generated pressure waves. The forward map, from the wave speed to the measurements, is nonlinear and oscillatory. The oscillations cause cycle skipping, the main impediment to using the standard, nonlinear least-squares data fitting formulation, known as full waveform inversion (FWI). A recently introduced alternative waveform inversion approach computes from the measurements an algebraic surrogate of the wave operator, a reduced order model (ROM) matrix, which is then used to estimate the wave speed. The mapping from the measurements to the ROM is nonlinear, but well understood. It is computed efficiently, in a non-iterative manner. The nonlinear mapping from the ROM to the wave speed is less understood, and its approximation involves time-consuming optimization. Our goal in this paper is to use a neural network to map the ROM matrix to a nearby one, that has a simpler and explicit dependence on the wave speed. This simplifies and reduces the computational cost of the ROM-based waveform inversion. We introduce the methodology, called ROMNet, and test it with numerical simulations, using two training data sets: The first set consists of random media with variations of the wave speed modeled by a superposition of Gaussians with random amplitudes and standard deviations. The second is the publicly available GeoFWI dataset introduced for benchmarking FWI using deep learning. We compare the performance of ROMNet with the direct ROM-based inversion and with two representative deep learning approaches to FWI: Fourier-DeepONet" andInversionNet”.

11. What Do Audio-Visual Synchronization Metrics Actually Measure?

Authors: Jai Kumar Sharma, Peeyush Tapadiya

Categories: cs.CV, cs.MM, cs.SD | Accepted at the ECCV 2026 Workshop on Generative AI for Audio-Visual Content Creation (Gen4AVC), poster presentation; non-archival workshop. 7 pages (4-page main text + references + 2-page appendix), 3 figures, 8 tables. Project page: https://jaishrm07.github.io/avsync-reliability-card/ Score: 6.21/10 (Obj:8 Id:6 Ind:5 Comp:6 Eff:8 Nov:6 Rep:3)

Strength: 四指标统一审计得出轴分裂结论——Synchformer 时域偏移 τ=0.84 而 PEAVS 代理仅 0.07，嵌入类指标相反，AV-Align 全面最弱（CV 0.54、PEAVS 0.00），且四指标互不同意（α=0.066），并给出可直接使用的 MDD 与 Reliability Card 报告指引。
Weakness: 人类对齐全靠 PEAVS 学习模型代理（非人类标签），主审计仅 75 clip，且审计代码与数据均未发布（README 明示 Code: Coming soon），标题承诺的感知真值未获直接验证。
Full review: Claude Code 全文七公理审稿

Automatic AV-sync metrics are widely used to rank and train audio-visual generators, but they are rarely audited as measurement instruments. We jointly audit AV-Align, ImageBind AV-relevance, JavisScore, and Synchformer/DeSync under a common reliability protocol: controlled-distortion monotonicity, preprocessing sensitivity, rank uncertainty, cross-metric agreement, PEAVS-proxy agreement, and learned fusion. The result is an axis split, not a single winner: Synchformer/DeSync is the strongest temporal-offset tracker ($τ=0.84$), ImageBind/JavisScore better match the PEAVS human-aligned proxy ($τ=0.20$) and content-disruption families, and AV-Align is the weakest standalone metric. The metrics mutually disagree (Krippendorff $α=0.066$), and neither linear nor simple $k$-NN fusion improves PEAVS agreement over the best individual metric. We recommend reporting AV-sync as a Reliability Card (metric-family breakdowns with confidence intervals) rather than a single bare synchronization score.

Authors: Hanyu Xuan, Mengqi Zhang, Junjun Mao, Fei Wang, Kun Li et al.

Categories: cs.MM Score: 6.11/10 (Obj:7 Id:10 Ind:6 Comp:5 Eff:6 Nov:5 Rep:3)

Strength: 任务解耦 LoRA 在六个音视频任务上全面超过开源基线 Crab，最显著增益为 RAVS unseen mIoU 45.6→52.1（+6.5），且与 Crab 共享 LLaMA-2-7B 基座与数据，对比同构公平。
Weakness: 多数任务增益仅为 +0.3~+1.7 点、AVQA 反降 0.3 点，方法为 HydraLoRA/MTL-LoRA/MoE 的既有组合，且全文无任何代码或权重发布声明，可复现性存疑。
Full review: Claude Code 全文七公理审稿

Inspired by human multi-modal perception, Audio-Visual Multi-Modal Learning (AVMML) integrates auditory and visual information to leverage complementary cross-modal cues, enabling more robust and comprehensive scene perception. Existing studies predominantly tackle each AVMML task in isolation, which stands in stark contrast to humans’ unified cognitive capacity for handling versatile perception. However, naive joint training across multiple AVMML tasks often suffers from mutual interference, arising from the intricate inter-task relationships. To address this, we propose a unified framework that simultaneously accommodates versatile AVMML tasks. Specifically, benefiting from powerful representation and generalization capabilities of large language models, we design a task-disentangled Low-Rank Adaptation (LoRA) mechanism that enables dynamic integration of both task-specific and task-shared knowledge, thereby facilitating effective multi-task collaboration. The proposed task-disentangled LoRA comprises three components: a task-general low-rank matrix, task-specific modulation matrices, and cross-task collaboration experts, which respectively capture universal audio-visual knowledge, decouple task-specific pattern, and exploit inherent inter-task correlations. By unifying explicit collaboration from both model and task perspectives, our approach not only surpasses existing unified audio-visual models across multiple AVMML tasks, but also outperforms most task-specific models on certain AVMML tasks.

13. Array-Agnostic Ambisonics Encoding via Diffusion Posterior Sampling

Authors: Amit Milstein, Nir Shlezinger, Boaz Rafaely

Categories: eess.AS, eess.SP Score: 6.10/10 (Obj:1 Id:2 Ind:1 Comp:1 Eff:2 Nov:2 Rep:1)

Strength: 训练一次扩散先验、推理期零样本适配任意阵列，理想场景 SI-SDR 全面领先线性（4-mic 11.66 vs 6.80 dB）与参数化基线，4-mic 上还胜过专门训练过的 U-Net。
Weakness: 代码仓库（Amitmils/AmbiDiffEnc）实测 404 不可访问，且与 8 天前 Moliner 等人的 device-agnostic diffusion Ambisonics 工作（2608.14097）高度重叠却未对比，失配场景 coherence 落后线性基线、推理成本零量化。
Full review: Claude Code 全文七公理审稿

Spatial audio enhances user immersion by reproducing 3D sound fields, with Ambisonics being a widely adopted representation. While Ambisonics is theoretically independent of the recording setup, practical microphone arrays introduce hardware-dependent encoding artifacts. Moreover, existing data-driven solutions lack flexibility, as they are typically restricted to fixed array geometries. To overcome these limitations, we propose ADEPS, a generative framework that explicitly embeds the physical acquisition model into the inference process. By leveraging this formulation, ADEPS effectively compensates for array-specific distortions while enabling zero-shot encoding across arbitrary array topologies. We train the underlying generative prior in an unsupervised manner solely on target Ambisonic representations. Extensive evaluations across diverse simulated and real microphone arrays demonstrate that ADEPS consistently outperforms both traditional linear and parametric baselines in spatial fidelity and spectral quality.

14. Benchmarking LLM Judges for Voice-Agent Evaluation: Reliability, Calibration, and Human Oversight

Authors: Anupam Purwar, Shashank Singh, Kritika Srivastava

Categories: cs.AI, cs.ET | Extends LLM-as-a-Judge to voice agents across telecom and retail, testing GPT-4.1 and GPT-5 against human raters across 10 safety and efficiency metrics. A correlation-based calibration analysis reveals domain-dependent reliability and identifies Recovery Turn Count and safety-recall metrics as unreliable for fully automated judging Score: 6.00/10 (Obj:1 Id:2 Ind:1 Comp:1 Eff:2 Nov:2 Rep:1)

Strength: 用 242 条零售/电信语音对话 × 10 指标 × GPT-4.1/GPT-5 × p0/p1/p2 三配置系统测量人-LLM 裁判分歧，发现 TE/CR 稳定（比率≈1）、安全指标 SR/IAS 在 Telecom 分歧达 3.5–6.1、RTC 最不稳定（1.889–6.064），且校准有效性随领域变化（Retail 跨指标 r=0.912–0.943 显著 vs Telecom GPT-4.1 r=0.295 不显著）。
Weakness: 论文无自身代码/数据发布，10 个指标与裁判提示完全依赖预印本 companion paper（其仓库 Data 目录仅有说明文件、无 242 条对话与标注），闭源裁判 API 采样设置零披露，”回算分数”方法未说明，线性校准只被提出而未被拟合验证。
Full review: Claude Code 全文七公理审稿

Evaluating conversational voice agents at scale re- quires reliable assessment methods that capture both observ- able interaction quality and the contextual judgment typically provided by human evaluators. We investigate LLM-as-a-Judge evaluation by comparing human judgments with GPT-4.1 and GPT-5 on telecom and retail voice-agent conversations, across conversational quality and safety dimensions. The same interac- tions are scored under three evaluation configurations, p0, p1, and p2, to test whether automated judgments are sensitive to the evaluation setup and whether observed patterns generalize across configurations and judge models. Beyond aggregate agreement, we examine metric-level correlations, evaluator consistency, and systematic human-LLM disagreement to identify which conver- sational attributes can be judged reliably by automation and which remain sensitive to interpretation and context. Effective voice-agent evaluation is also shaped by pipeline-level factors such as speech generation, streaming, and error propagation across ASR, reasoning, and tool-calling stages, motivating our focus on comparing how human and LLM judges score the same interactions end to end. Our results show that LLM- based evaluation can serve as an effective component of large- scale voice-agent assessment, but that its reliability is metric- and configuration-dependent rather than uniform. This pro- vides an empirical framework for identifying which metrics suit automated evaluation and supports hybrid pipelines in which LLM judges handle scalable assessment while human evaluators remain engaged for metrics that demand contextual interpretation and higher-confidence judgment.

15. Investigating voiced and unvoiced regions of speech for audio deepfake detection

Authors: Ganesh Sivaraman, Hemlata Tak, Elie Khoury

Categories: eess.AS, eess.SP | Accepted in IEEE ICASSP 2025 Score: 5.95/10 (Obj:1 Id:2 Ind:1 Comp:1 Eff:2 Nov:2 Rep:1)

Strength: 清音区仅占音频时长约 27%，却以 6.62% EER 优于全音频基线 11.40%，清浊分数融合降至 5.82%，相对全音频基线降低 49%。
Weakness: 自训 AASIST 基线（11.40% EER）远弱于该模型在 ASVspoof2019LA 上的约 1% 水平，评测仅限 MLAAD 单语料、无跨语料验证，且代码与权重未公开。
Full review: Claude Code 全文七公理审稿

Deep neural network based deepfake detection systems have achieved high levels of accuracy on benchmark datasets and competitions. However, most models lack interpretability. It is challenging to extract reasoning from the network that can convince the human evaluator to trust the decision. Humans often rely on acoustic cues like unnatural pitch jitter, robotic intonation, acoustic artifacts, and unnatural sounding fricatives to judge the quality of the synthetic audio. This study explores the role played by the voiced and unvoiced regions of speech in discriminating synthetic from bonafide speech. A measure of signal periodicity is used to analyze speech into voiced and unvoiced components. Then, the graph attention based AASIST detection system is trained independently on each component. This work compares the accuracy of deepfake detection system using voiced and unvoiced components and analyzes the results on the MLAAD dataset. Our results show that unvoiced regions are particularly more effective in distinguishing synthetic (deepfake) speech from bonafide, and achieves an equal error rate of 6.62%. When combined with voice regions through score-level fusion, the overall performance improves further, yielding a 5.82% EER, a relative improvement of 49% over the baseline system that uses the full audio.

16. Visually-Guided Spatial Audio Generation for $360^\circ$ In-the-Wild Speech Scenes

Authors: Qingyu Luo, Peng Zhang, Wenwu Wang, Philip J. B. Jackson

Categories: eess.AS | Accepted at INTERSPEECH 2026 Score: 5.79/10 (Obj:1 Id:2 Ind:1 Comp:1 Eff:2 Nov:2 Rep:1)

Strength: 提出首个语音导向的野外 360° 视频-FOA 数据集 YT-SPEECH（8.9h、197 视频），Localizer-Renderer 完整模型在 Δang（0.64°）与 PESQ（3.42）上优于全部消融基线。
Weakness: 代码与数据集均未发布，评测指标与训练损失同源，9 人小样本主观评测且无外部第三方对比，相对 Localizer-AmbiEnc 在 MRS/MOS-Q 上反而落后。
Full review: Claude Code 全文七公理审稿

Spatial audio is a key component of immersive $360^\circ$ media, yet high-quality spatial capture remains limited in real-world speech-dominant scenes. We study visually guided First-Order Ambisonics (FOA) speech spatialization in the wild: given aligned $360^\circ$ video and an omnidirectional audio track, we recover the missing directional FOA components. To support this task, we introduce YT-SPEECH, a speech-oriented $360^\circ$ video-FOA dataset curated from YouTube. We propose a two-stage Localizer-Renderer framework, where an audio-visual segmentation backbone provides frame-wise spatial heatmaps and a conditional complex-domain U-Net reconstructs directional FOA signals from the omnidirectional channel. A confidence-based gating strategy stabilizes conditioning under ambiguous acoustic conditions. Experiments show improved reconstruction fidelity, spatial accuracy, and perceptual speech quality relative to ablated variants and prior approaches.

17. ‘Ghaib in Translation’ aka Unseen Harm: Measuring Cross-Script Safety Inconsistency with ‘Missed-in-Urdu’ Scores in LLM Hate Speech Detection

Authors: Fawzia Zehra, Kara-Isitt, Sonal Khosla, Stephen Swift

Categories: cs.CL, cs.AI Score: 5.53/10 (Obj:6 Id:5 Ind:9 Comp:8 Eff:3 Nov:4 Rep:8)

Strength: 五年 WOAH/ALW 共 205 篇文献审计显示乌尔都零覆盖（论文 §1 + 附录），并以五个商业+开源 LLM 在五乌尔都数据集上的”Missed-in-Urdu”中位数 4.3%（2.4–9.9%）实证 LLM 跨文字安全不一致性，仓库 FuzzyLogic9/WOAHJul26 公开 5 个 Python 脚本 + 9 MB 原始 CSV + 4 张 figure。
Weakness: 未与 MultiJail、Mu-SHROOM、PURE 等现成跨语种安全基准做 head-to-head；”Missed-in-Urdu”只算单边指标（放行乌尔都有害），缺少对称的”错杀乌尔都正常”率；闭源 LLM 温度=0 单次采样无置信区间；论文止于描述不提出 mitigation，效用对工程团队接近零；基线论文检索因本机 arXiv 路由不可达（DNS 工作区限制）无法独立验证。
Full review: Claude Code 全文七公理审稿

Urdu, the world’s tenth most spoken language with 246 million speakers, remains almost entirely absent from mainstream LLM safety evaluation and nine years of WOAH proceedings. To investigate whether this absence has measurable consequences for content moderation reliability, five large language models, GPT-4o, Claude Sonnet 4.5, Gemini 2.5 Flash, Qwen-2.5, and Llama-3.1, were tested across six datasets spanning Nastaliq Urdu, Roman Urdu, English, and code-switched Urdu-English. Across the five Urdu-script datasets, label instability between original-script and English-translation classification ranged from 15.9% (Gemini 2.5 Flash) to 31.6% (Qwen-2.5), with a ‘Missed-in-Urdu’ rate, content flagged as harmful in English translation but passed as normal in the original script, ranging from 2.4% to 9.9% (median 4.3%). A complete enumeration of all 205 papers across nine ALW/WOAH editions via the ACL Anthology API confirms zero dedicated Urdu papers across the entire period. Results indicate that current LLMs provide uneven safety assurance across Urdu’s script varieties, with smaller open-weight models showing substantially higher instability and missed-harm rates than frontier closed models.

18. Don’t Just Listen, Try Planning: Graph-based Retrieval-Generation Agent for Long-form Audio Meeting Understanding

Authors: Quanwei Tang, Dong Zhang, Shoushan Li, Guodong Zhou

Categories: cs.SD, cs.AI | ACL Findings 2026 Accepted Score: 5.42/10 (Obj:8 Id:5 Ind:5 Comp:4 Eff:6 Nov:5 Rep:5)

Strength: 自建 16,456 条 LongAudioQA 长音频会议 QA 数据集，GRGA 在五类任务上全面领先最强 Speech LLM 基线（AMI Infer 65.29 vs MiMo-Audio 44.10；Acoustic 35.68 vs 17.17），High/High 噪声下仍保 39.54%、对 Text RAG 优势扩至 +14.87%。
Weakness: 数据、图、作答三端同源且证据标注与检索目标以 IoU>0.9 对齐，评测独立性弱；基线全部单轮检索，缺失最强替代对照；代码仓库 URL 缺失、GitHub 0 命中，可复现承诺无法核验。
Full review: Claude Code 全文七公理审稿

While long-form audio meeting understanding (LAMU) is garnering growing attention, task-specific question answering (QA) datasets remain scarce. Existing speech QA paradigms and state-of-the-art Speech LLMs suffer from acoustic information loss and poor long-term context memory. To address these issues, we construct the LongAudioQA dataset and propose the GRGA model, which models heterogeneous audio features into a multi-dimensional graph and leverages agent planning for retrieval and answer generation.

19. AudioLens: Multi-Perspective Speech Clustering with Reasoning Audio-Language Models

Authors: Wenjun Huang, Qiaosong Chu, Tiger Shao, Pengfei Zhang, Yutong Song et al.

Categories: cs.SD, cs.AI Score: 5.32/10 (Obj:8 Id:7 Ind:5 Comp:5 Eff:5 Nov:5 Rep:2)

Strength: AudioLens-R1 在自建 AudioLens-Bench 上总体 ARI 44.77（V-measure 73.43），较最强基线 GPT-audio-1.5 提升 12.99/11.62，情绪视角 ARI 从 2.40 升至 39.00。
Weakness: 基准全为 TTS 合成语音且代码/数据/权重均未发布，模型侧为推理蒸馏+DPO 的既有组合，无自然语音与延迟/成本验证。
Full review: Claude Code 全文七公理审稿

Audio clustering is a fundamental task for organizing rapidly growing speech collections, supporting applications such as conversational analysis and speech-driven discovery. However, existing methods rely on fixed acoustic similarity metrics or ASR-based text pipelines, limiting their ability to reorganize the same audio collection under different user-specified perspectives, especially when clustering depends on both linguistic and paralinguistic cues. We introduce audio multi-perspective clustering, where a model directly partitions speech recordings according to a natural-language perspective while inferring both the number of clusters and their assignments. To study this setting, we construct AudioLens-Bench, a benchmark spanning multiple application domains and evaluating both in-perspective and cross-perspective generalization. We further propose AudioLens-R1, an end-to-end large audio-language model trained with reasoning distillation and preference optimization. Experiments show that AudioLens-R1 consistently outperforms all baselines, improving overall ARI by 12.99 points and V-measure by 11.62 points. These results demonstrate the promise of native audio-language models for flexible, perspective-conditioned structure discovery over speech collections.

20. CoSTALA: Compositional Spatio-Temporal Audio-Language Alignment via Multi-Grain Hierarchical Contrastive Learning

Authors: Peiwei Ren, Jinbo Hu, Fang Kang, Shan Liang, Yin Cao

Categories: cs.SD Score: 5.11/10 (Obj:1 Id:2 Ind:1 Comp:1 Eff:2 Nov:2 Rep:1)

Strength: 提出四重多粒度分层对比损失，完整开源代码与数据合成管线，消融覆盖全部损失组合，完整模型 Global 检索 R@1 达 8.10%（T2A）。
Weakness: 基线数字错误引用（SALM sClotho R@1 实为 10.5% 却被记为 4.77%），评测仅限合成双事件数据，核心优势建立在被低估的基线上。
Full review: Claude Code 全文七公理审稿

Conventional audio language models (ALMs) have made significant progress in achieving alignment between auditory and textual representations, including recent explorations in spatial audio. However, in daily spatial scenarios, they still cannot effectively process multi-event audio sequences. Current approaches primarily rely on coarse-grained contrastive learning with global auditory and textual features, lacking the resolution to distinguish multiple sequential events. To overcome these limitations, we propose CoSTALA-a novel training paradigm that transitions from purely global alignment to fine-grained spatio-temporal reasoning. By constructing a multi-granularity hierarchical loss function system, we achieve explicit modeling of temporal dependencies, and successfully anchors individual acoustic events to preserve their semantic purity. Extensive experiments demonstrate that CoSTALA significantly establish a powerful new framework for spatio-temporal audio understanding.

21. On the Robustness of Audio Deepfake Detection under Audio Watermarking

Authors: Zi Qian Yong, Ajinkya Kulkarni, Julia Lau, Hwa Hui Tew, Shu Min Leong et al.

Categories: cs.SD Score: 5.10/10 (Obj:1 Id:2 Ind:1 Comp:1 Eff:2 Nov:2 Rep:1)

Strength: 黑盒评测 8 个 ADD 模型 × 5 数据集，WavMark 水印（SNR 38.55 dB）在 ASVspoof 2021 LA/DF 上使 EER 从约 2% 恶化至 29–35%，且嵌入位移（FD 最高 1089.69）与检测退化强相关，实证数据完整。
Weakness: 缺等强度噪声/失真对照而无法归因于水印、仅测单一水印、且摘要承诺的代码仓库为空占位，均削弱结论强度与可复现性。
Full review: Claude Code 全文七公理审稿

Recent advances in generative audio models have enabled highly realistic synthetic speech, increasing the importance of reliable audio deepfake detection (ADD) systems. While prior studies have primarily focused on adversarially optimized perturbations, the robustness of ADD systems under realistic signal transformations remains insufficiently understood. In this work, we investigate the impact of audio watermarking on ADD systems by treating watermarking as a structured, non-adversarial perturbation rather than a conventional attack mechanism. Using a watermark-based evaluation framework built upon WavMark, we evaluate multiple self-supervised learning (SSL), Convolutional Neural Network (CNN) and Graph Neural Netrowk (GNN)-based ADD models across several benchmark datasets. Beyond conventional detection metrics, we further analyze watermark-induced representation shifts using Fréchet Distance, cosine similarity, and L2 distance in the embedding space. Experimental results reveal a strong dataset-dependent behavior: watermarking causes substantial performance degradation on ASVspoof 2021 LA and DF, while exhibiting limited impact on ASVspoof 2024, FoR, and ITW. Moreover, large embedding-space shifts are strongly associated with severe detection degradation, suggesting that watermark-induced perturbations can substantially alter the feature representations relied upon by current ADD systems. These findings demonstrate that benign signal transformations designed for content protection can expose previously overlooked robustness vulnerabilities in audio deepfake detection systems. Our code is available at https://github.com/ziqian0925/wm-ADD-robustness.git