Daily Papers — 2026-07-01

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

1. AV-SyncBench: Decoupled Benchmarking of Temporal and Semantic Audio-Visual Synchronization

Authors: Tianhong Zhou, Mingyang Han, Boyu Li, Yuxuan Jiang, Jiaxin Ye et al.

Categories: cs.CV, cs.SD

| Accepted by Interspeech 2026 Score: 7.05/10 (Obj:8 Id:5 Ind:6 Comp:8 Eff:6 Nov:8 Rep:8)

Strength: 中 — 解耦构造（时间扰动/语义编辑保持正交）是一个干净的 reframing，可被后续工作复用；语义挑战任务用生成式编辑保时序是可借鉴的构造范式。
Weakness: Benchmark 论文的”识别”=能否准确诊断模型在哪一维度强/弱。解耦设计本身做到了变量隔离（时间任务只动时序、语义任务只动语义），这一点是合理的。
Full review: Claude Code 全文七公理审稿

中文摘要： 音视频特征提取是多模态理解与生成任务的基础，但现有评估协议存在维度偏差，通常只侧重语义匹配或时序偏移检测中的某一项，且数据构建方式相互耦合，无法对时序一致性与语义一致性进行独立评估。针对这一不足，作者提出AV-SyncBench，这是首个将时序评估与语义评估完全解耦的基准测试框架，通过分离式数据构造实现两个维度的独立量化。该基准填补了音视频同步评测中长期缺失的解耦评价能力，为多模态特征提取模型的精细诊断与针对性改进提供了更可靠的评价工具。

Audio-visual feature extraction is a fundamental component of multimodal understanding and generation tasks. However, existing evaluation protocols for feature extraction models exhibit dimensional bias, typically focusing on either semantic matching or temporal offset detection. Moreover, their data construction remains coupled, preventing independent assessment of temporal and semantic consistency. We propose AV-SyncBench, the first benchmark to fully separate temporal and semantic evaluation

2. What’s a Credit Worth? A Market Framework for Attribution-Aware Compensation in Generative Music

Authors: Luyang Zhang, Xirui Jiang, Junwei Deng, Beibei Li, Jiaqi W. Ma et al.

Categories: cs.CY, cs.LG Score: 7.05/10 (Obj:9 Id:6 Ind:8 Comp:8 Eff:6 Nov:8 Rep:6)

Strength: 首次将ML数据归因噪声（信号信息量Ij）作为合同设计的一阶输入，推导出闭式支付规则ρ=clip(Ij/αj,[0,1])，并揭示了”当前归因方法仅关闭约10%福利差距、22个creator中最多1个跨过版税阈值”这一重要且可靠的定量诊断。
Weakness: 实证部分缺乏最简baseline对比和与前期工作(Deng et al. 2023)的直接定量对比，且per-creator bootstrap CI多数跨零（T=200样本量偏小），无法在个体层面验证理论的核心预测。
Full review: Claude Code 全文七公理审稿

中文摘要： 这篇论文聚焦生成式音乐平台中创作者补偿的核心问题：当创作者的录音被用于训练生成式AI模型，而模型产出的音乐又具有商业价值时，应如何公平地向创作者支付报酬。作者提出了一个归因感知的补偿市场框架，其中每位创作者的报酬由数据归因评分决定，该评分衡量其作品对模型生成输出的贡献程度。框架将归因评分与市场定价机制相结合，使补偿能够反映每位创作者数据对生成结果的边际价值。该方法为生成式音乐市场的平台设计提供了可操作的分配方案，在激励创作者持续贡献数据与维持平台商业可持续性之间寻求平衡。

Advances in generative AI are rapidly increasing the quality and commercial value of generated music, and this progress depends on large catalogs of creators’ recordings. This raises a central question for platform design: how should creators be compensated when their work is used to train generative AI models that in turn produce commercial outputs? We develop a framework for fairly compensating creators in generative-music markets, where each creator’s payment depends on a data-attribution sco

3. Evaluating Pretrained Music Embeddings for Cross-Performance Jazz Standard Recognition

Authors: Çağrı Eser

Categories: cs.SD, cs.LG

| 6 pages, 2 figures, 4 tables. Accepted to the ICML 2026 Workshop on Machine Learning for Audio Score: 6.58/10 (Obj:8 Id:5 Ind:8 Comp:8 Eff:5 Nov:6 Rep:8)

Strength: 低到中 — 无新架构、新损失、新训练算法；SupCon 投影是现成组件的应用层组合。方法本身不构成可被他人复用的工具性贡献。
Weakness: (1) 缺 cover-song ID 强 baseline（ByteCover/CoverHunter/DisCover）和经典 chroma+DTW baseline，使”预训练嵌入比从零训练好”的结论缺乏最强对照；(2) SupCon 消融不干净，把”加投影+加 CE+加 SupCon”三件事的提升归因于 SupCon；(3) 绝对 Top-1 仅 12.5%，论文未给人类天花板锚定，使”任务在当前方法下基本不可用”这一事实的解读。
Full review: Claude Code 全文七公理审稿

中文摘要： 本文研究跨表演的爵士标准曲识别这一tune-level音乐检索任务，难点在于同一标准曲的不同表演在速度、调性、编排、乐器和即兴内容上可能差异巨大，甚至head旋律完全不出现。作者基于Jazz Trio Database构建 curated 子集（16 standards / 79 performances，leave-one-performance-out 协议），将从零训练的 Harmonic CNN 与冻结的预训练音乐嵌入（MERT、MuQ 等）在统一协议下进行 probing 与 retrieval 对比。主要发现是预训练嵌入整体优于从零训练基线，但绝对效果仍很低（最佳 Top-1 仅 12.5%，16 选 1），且嵌入空间编码表演者身份强于编码 tune 身份，导致检索被 performer bias 主导。作者进一步提出用轻量 SupCon 投影显式拉近同 standard 跨表演样本，能在 MERT 上将 Same Group Freq 从 0.336 降至 0.109，但对 Top-1 的提升在不同 encoder 上不一致。论文将该任务定位为音乐 foundation 模型的 exploratory stress test，并诚实报告了窗口长度、PCA 降维等多组负面发现。

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 m

4. Positive-Incentive Noise Predictor for Adversarial Purification in Speaker Verification

Authors: Yibo Bai, Sizhou Chen, Michele Panariello, Hao Ma, Xiao-Lei Zhang et al.

Categories: eess.AS

| Submitted to IEEE TASLP.13 pages for maunscript, 2 pages for supplementary material Score: 6.30/10 (Obj:8 Id:6 Ind:8 Comp:8 Eff:6 Nov:5 Rep:8)

Strength: 发现扩散净化中前向加噪贡献绝大部分鲁棒性，据此提出 PnP-Diff 将净化简化为单步可学习加噪，RTF 降至 0.014（较 AudioPure 快 3.6 倍），白盒 pooled EER 达 3.02%，并在四种 ASV 后端上验证了跨架构迁移性。
Weakness: 与最简 baseline Noise-0.01（无需训练、pooled EER 3.40%）的性能差距仅 0.02-0.58 个百分点，增量收益相对复杂度不明显；核心”前向主导”观察在图像域已有先例（Liu & Wei, ICLR 2024）未被引用；与 SSNI (ICML 2025) 的思路高度相似。
Full review: Claude Code 全文七公理审稿

中文摘要： 针对自动说话人确认（ASV）系统易受对抗扰动攻击、而基于扩散的净化方法又因反向去噪的迭代采样导致推理延迟过高的问题，作者发现扩散净化中前向加噪过程其实贡献了大部分鲁棒性增益，据此将对抗净化重新表述为一个可学习的加噪问题，并提出首个显式引入”正向激励噪声”（π-noise）的框架 PnP（Positive-Incentive Noise Predictor），通过一维 U-Net 学习输入自适应的 π-noise 并将其与原始输入混合，以提升下游 ASV 系统的鲁棒性。在四种先进 ASV 主干网络上的实验表明，PnP 既能有效防御对抗攻击，又能保持自然语音上的识别性能，在白盒、黑盒及防御者感知的自适应攻击下，于防御效果、对真实语音的影响与推理效率之间取得了有竞争力的平衡，实时因子最低可达 0.014，远优于扩散净化的效率；此外，PnP 还可与扩散去噪器级联，进一步提升净化语音的感知质量。

Modern automatic speaker verification (ASV) systems are vulnerable to adversarial perturbations. Diffusion-based purification has recently shown strong effectiveness against such perturbations, but its reverse denoising process requires iterative sampling and leads to high inference latency. We find that the forward noising process provides most of the robustness gain. Motivated by this observation, we reformulate adversarial purification as a learnable noising problem, and propose the Positive-

5. From Objectives to Applications: Aligning Architectural Biases in Audio Self-Supervised Learning

Authors: Kele Xu, Yulu Fang, Boda Zhou, Yulin Sun, Qisheng Xu et al.

Categories: eess.AS Score: 6.26/10 (Obj:8 Id:5 Ind:5 Comp:7 Eff:7 Nov:6 Rep:6)

Strength: 低到中 — 综述无新方法；分类法对新人有导航价值，对资深研究者增量为中等。
Weakness: 综述的”识别”=对方法机制给出准确因果归因。论文提出”目标→表征需求→架构偏置”的因果链（如：对比学习→不变性→CNN 局部偏置；masked 重建→上下文推断→Transformer 全局路由；离散 token→混合架构；多模态→跨模投影）。这些归因是事后诠释性叙事，并非通过受控消融或定量验证得到。论文同时承认 teacher-student 是”cross-cutting mechanism”，分类边界本身有重叠。
Full review: Claude Code 全文七公理审稿

中文摘要： 本文从预训练目标、架构归纳偏置与下游应用三者之间的对齐关系出发，重新审视音频自监督学习（SSL），而非将SSL方法简单视为按时间顺序排列的预训练任务或模型族系。核心问题在于：不同的监督信号如何塑造模型所期望学习的表征，以及这些表征是否与下游任务的需求相匹配。作者围绕辅助任务、对比学习、生成式重建、离散token化等五种范式组织讨论，系统梳理每种范式下目标函数与架构选择之间的相互作用。该方法论框架有助于解释为何某些SSL方法在特定下游任务上表现优异而在其他任务上失效，并为根据应用场景选择合适的预训练策略提供了指导。主要贡献在于提供了一个统一的分析视角，使研究者在设计音频SSL系统时能够更有意识地对齐目标、架构与应用三者之间的关系。

This paper examines audio self-supervised learning (SSL) through the alignment between pretraining objectives, architectural inductive biases, and downstream applications. Rather than treating SSL methods as a chronological sequence of pretext tasks or model families, we ask how different supervisory signals shape the representations that models are expected to learn. The discussion is organized around five paradigms: auxiliary tasks, contrastive learning, generative reconstruction, discrete tok

6. NPUsper: Eliminating Redundant Computation for Real-Time Whisper on Mobile NPUs

Authors: Sihyeon Lee, Hojeong Lee, Sungwon Woo, Chengpo Yan, Suman Banerjee et al.

Categories: cs.SD Score: 6.11/10 (Obj:8 Id:6 Ind:8 Comp:6 Eff:6 Nov:5 Rep:8)

Strength: 通过交叉注意力 backward shift 在线检测幻觉 token 以消除 padding overhead，结合 K 步 chunk graph 的 controlled unrolling 适配 NPU 静态图执行，在 Samsung Galaxy S25 上实现 4.84x 更低的每词延迟和 88.64% 更低的平均功耗。
Weakness: WER 在 TED-LIUM 3 上比 SimulStreaming/Simul-Whisper 高约 3pp（14.18% vs. 11.26%/11.40%），论文声称 “comparable accuracy” 不够诚实；且同期 on-device Whisper 系统 WhisperKit (2507.10860) 未被引用或比较。
Full review: Claude Code 全文七公理审稿

中文摘要： 针对Whisper语音模型在移动NPU上实时转写时存在大量冗余计算的问题，作者提出了NPUsper系统，通过消除冗余计算来提升效率。为避免以往流式系统所依赖的大量padding，NPUsper利用decoder cross-attention中的时间模式在线检测幻觉token，使每轮推理仅需处理短音频输入并最小化carryover。针对移动NPU的执行特点，系统提出controlled unrolling技术，将自回归解码以K步chunk图的形式高效执行。该系统实现了Whisper在移动设备上的实时高效转写。

We present NPUsper, a live transcription system that makes Whisper efficient on mobile NPUs by eliminating redundant computation. To avoid the heavy padding used by prior streaming systems, NPUsper detects hallucinated tokens online from temporal patterns in decoder cross-attention, allowing each inference round to process short audio inputs with minimal carryover. For efficient mobile-NPU execution, we propose controlled unrolling, which executes autoregressive decoding as K-step chunk graphs,

7. Few-Shot Open-Set Audio Classification Using Attention Information-Fused Prototypes

Authors: Yanxiong Li, Jiaxin Tan, Qianqian Li, Guoqing Chen, Sen Huang et al.

Categories: eess.AS, cs.LG

| 14 pages, 12 tables, 9 figures,Accepted for publication in IEEE TASLP Score: 6.05/10 (Obj:6 Id:5 Ind:8 Comp:4 Eff:7 Nov:5 Rep:8)

Strength: 在 3 个公开数据集 + 1 个额外数据集 + 6 组跨数据集实验上全面优于 8 种基线方法（Acc 和 AUROC），统计显著性强于除 MET/D-ProtoNet 外的所有基线，且计算复杂度 MACs/AIT 排名前列，代码和数据完全开源。
Weakness: 核心组件（attention+prototype 融合、MLP 负原型、attention module）均有明确已有来源且第一作者自身前期工作已探索类似组件，缺少最简阈值基线对比和子模块级消融，FSC-89 绝对 AUROC 仅 71% 且部分提升幅度 <1%。
Full review: Claude Code 全文七公理审稿

中文摘要： 现有音频分类方法大多假设测试样本必属于训练类别，无法识别并拒绝未知类样本。本文针对少样本开集音频分类（FOAC）问题提出一种新方法，能在仅用少量支持样本更新模型后正确识别已知类查询样本，同时拒绝来自未知类的查询样本。该方法设计了基于注意力信息融合的原型表示模型，通过将注意力机制与原型学习相结合来增强类别特征的判别能力，从而更好地区分已知类与未知类。实验表明该方法在保持已知类识别精度的同时，显著提升了未知类样本的拒绝能力，为开放场景下的少样本音频分类提供了有效解决方案。

Most existing audio classification methods suppose that each query (testing) sample belongs to a class of support (training) samples, and misrecognize samples of unseen classes as seen classes (cannot reject samples of unseen classes). In this study, we propose a method for Few-shot Open-set Audio Classification (FOAC), which can recognize query samples of seen classes after updating the model using a few support samples, and meanwhile reject query samples from unseen classes. We design a model

8. TurnNat: Automatic Evaluation of Turn-Taking Naturalness in Dyadic Spoken Dialogue

Authors: Hao Zhang, Thomas Thebaud, Georgi Tinchev, Venkatesh Ravichandran, Laureano Moro-Velazquez

Categories: cs.CL, cs.AI Score: 5.95/10 (Obj:6 Id:5 Ind:5 Comp:8 Eff:5 Nov:5 Rep:8)

Strength: 中 — VAP 训练目标 + DualTurn backbone + NLL 聚合不是新方法，是合理组合；消融合格但缺最简 baseline 与外部 baseline。
Weakness: 识别设计存在真实缺陷。论文做了一系列消融（V0 vs V1 vs D0-D4，TBU 权重 α=1/3/8，categorical vs Bernoulli output，aux supervision），这部分是合格的。
Full review: Claude Code 全文七公理审稿

中文摘要： 这篇论文聚焦于全双工口语对话系统中轮流发言自然性的自动评估问题，指出已有方法多依赖人工评判或针对单一时序行为的指标，难以在统一框架下比较异构的时序失败。作者提出TurnNat，一种面向双通道口语对话的基于似然的自动评估框架，通过在自然对话数据上训练因果轮流发言预测模型来刻画自然轮流模式。该框架将不同类型的时序失误纳入统一度量，从而支持对不同系统在轮流自然性上的横向比较。实验表明TurnNat与人类自然性评判具有较高一致性，并能敏感区分各类典型时序错误，为口语对话系统的轮流质量评估提供了可扩展的自动化方案。

Turn-taking naturalness is central to full-duplex spoken dialogue systems, yet its automatic evaluation remains limited. Existing evaluations often rely on human judgments or behavior-specific timing metrics, making it difficult to compare heterogeneous timing failures within a unified framework. We propose TurnNat, a likelihood-based framework for automatic turn-taking naturalness evaluation in two-channel spoken dialogue. A causal turn-taking prediction model trained on natural conversations e

9. Enhancing Flow Matching with A Unified Guidance Framework for Efficient and Robust Speech Synthesis

Authors: Zuda Yu, Qianhui Xu, Ting Chen, Junhui Zhang, Tao Fu et al.

Categories: cs.SD, cs.AI

| Accepted to INTERSPEECH 2026 Score: 5.53/10 (Obj:7 Id:5 Ind:6 Comp:5 Eff:6 Nov:5 Rep:5)

Strength: DG+MG 组合在工程上确实把 3-NFE 下 SIM 从 Enhanced MG-only 的 0.792 拉回 0.850，3.25× 加速与 SIM 双赢的工程结果是实的，可作为 INTERSPEECH 系统级贡献。
Weakness: 组件级消融存在（DG-only / Vanilla MG / Enhanced MG / Unified），DG 与 MG 各自贡献可识别。
Full review: Claude Code 全文七公理审稿

中文摘要： 该论文针对Flow Matching在语音合成中存在的推理延迟高和音色泄露问题，提出了一个统一指导框架，通过数据指导和模型指导两种互补策略提升生成效率与鲁棒性。数据层面采用异构增强技术，促使模型将语言内容与声学残差解耦；模型层面则通过改进的Mo（原文截断）策略进一步增强性能。该框架的核心在于同时从数据和模型两个维度入手，协同解决FM语音合成的关键瓶颈。

Flow Matching (FM) has emerged as a powerful paradigm for speech generation but remains constrained by high inference latency and timbre leakage. To address these bottlenecks, we propose a unified guidance framework that enhances generation efficiency and robustness through two complementary strategies. On the data front, we introduce Data-guidance via heterogeneous augmentation, encouraging the model to disentangle linguistic content from acoustic residue. In parallel, we propose an enhanced Mo

10. A Geometric Perspective on Composable Emotion Steering in Text-to-Speech Models

Authors: Siyi Wang, James Bailey, Ting Dang

Categories: cs.SD, cs.LG Score: 5.50/10 (Obj:8 Id:5 Ind:6 Comp:8 Eff:5 Nov:5 Rep:5)

Strength: 通过线性探测和LID分析首次揭示SLM与CFM在情感表示几何上的结构差异（SLM正ΔLID=+0.84表示情感子空间独立，CFM负ΔLID=-1.48表示情感-说话人纠缠），并实验验证SLM steering在比例控制上优于CFM（ρ=0.209 vs 0.193 on CREMA-D）。
Weakness: 所有实验仅在CosyVoice2单一架构上完成且缺少随机方向baseline隔离几何因果解释，代码未开源，核心发现”SLM更适合emotion steering”与前序工作CoCoEmo（同作者组）的实验结论高度重叠。
Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对文本到语音（TTS）系统中情感控制模块的几何性质及其对可调控性影响尚不明确的问题，开展了首项针对语音语言模型（SLM）与条件流匹配（CFM）模块作为激活调控位点的对比研究。作者首先利用线性探测与局部内在维度（LID）刻画这两种模块中的情感表示几何结构，随后在混合情感语音合成任务上评估其作为激活调控位点的可行性与效果。该研究将表示几何性质与可调控性联系起来，为理解不同TTS模块在组合情感控制中的差异提供了系统性的实证视角，并为如何在混合情感场景中选择合适的调控位点提供了指导。

While prior work has explored emotion control in hybrid text-to-speech systems, the geometric properties of these modules, and their implications for steerability, remain poorly understood. We present the first comparative study of speech language model (SLM) and conditional flow-matching (CFM) modules as activation steering sites for mixed emotion speech synthesis. We first characterize emotion representations using linear probing and local intrinsic dimensionality (LID), and then evaluate sing

11. Disentangling Speaker and Language Effects in Cross-Lingual Speaker Verification for Iberian Languages

Authors: Pol Buitrago, Javier Hernando

Categories: eess.AS, cs.CL

| 5 pages, 8 figures, Submitted to IberSPEECH 2026 Score: 5.42/10 (Obj:6 Id:5 Ind:9 Comp:4 Eff:5 Nov:2 Rep:9)

Strength: 无 — 没有提出新方法、新模型或新表示；CLTM 与 bilingual 协议均沿用既有工作。
Weakness: 论文真实增量非常薄：(1) CLTM 框架来自同作者 ref [6] 的 Interspeech 2026 论文，本论文直接套用；(2) “same-speaker bilingual evaluation” 不是新想法——ref [1] Winters et al. 2008、ref [11] Misra & Hansen 2014、ref [14] Nam et al. 2023 都已用双语说话人评测 cross-lingual S。
Full review: Claude Code 全文七公理审稿

中文摘要： 该论文针对跨语言说话人验证中语言失配与说话人间差异相互混淆的问题，构建了一个覆盖五种伊比利亚语言的双语同说话人评测集，从而在保持说话人身份不变的前提下分离语言效应与说话人效应。基于此评测集，作者系统分析了语言失配对验证性能的独立影响，揭示了传统评测协议因混淆两种效应而高估语言失配损害的程度。研究还比较了不同语言对之间的失配影响差异，为理解和缓解跨语言说话人验证的性能退化提供了新的分析视角与基准数据。

Cross-lingual speaker verification (SV) systems typically exhibit performance degradation when enrollment and test utterances are spoken in different languages. However, standard evaluation protocols confound language mismatch with inter-speaker variability, as evaluation is generally performed with different speakers across languages. In this work, we introduce a bilingual same-speaker evaluation set for five Iberian languages, enabling analysis of cross-lingual SV under constant speaker identi

12. From Monolingual to Multilingual: Evaluating Mamba for ASR in South African Languages

Authors: Jesujoba O. Alabi, Julian Herreilers, Badr M. Abdullah, Dietrich Klakow

Categories: cs.CL

| under review Score: 5.37/10 (Obj:6 Id:5 Ind:8 Comp:5 Eff:5 Nov:3 Rep:6)

Strength: 低 — 无新架构、无新训练方法、无新机制；五个多语言策略均为标准技巧组合。
Weakness: 组件全部已有：Mamba (Gu&Dao), ConMamba (Jiang Speech Slytherin [30]), CTC (Graves), language embedding as bias (LAE Tian 2022 [31]), language-family embedding, multitask LID。
Full review: Claude Code 全文七公理审稿

中文摘要： 这篇论文针对南非语言自动语音识别（ASR）领域研究不足的问题，系统评估了新型状态空间模型 Mamba 在七种南非语言上的表现。研究者在单语和跨语言两种设置下，对每种语言使用 50 小时语音数据进行训练，并将 Mamba 与 CTC 和 Conformer 等基线模型进行对比。通过在不同语言间的迁移实验，作者分析了 Mamba 在非洲低资源语言上的适应能力与性能特征。该工作为南非多语言 ASR 系统的架构选择提供了实证依据，揭示了 Mamba 相对于传统 Conformer 模型在非洲语言场景下的优势与局限。

Recent advances in automatic speech recognition (ASR) have explored different sequence models, including Conformer-based models and newer state space models such as Mamba. Although prior work has evaluated these architectures in multiple languages, their effectiveness in African languages remains underexplored. In this work, we evaluate Mamba for ASR on seven South African languages. In monolingual experiments, each model is trained on 50 hours of speech per language, and we compare Mamba to a C

13. A Text-Steerable Instrument for Sketching Procedural Soundscapes via Language Models

Authors: Prabal Gupta

Categories: cs.SD, cs.CL, cs.HC, eess.AS

| 10 pages, 7 figures, 2 tables. Accepted to the International Conference on New Interfaces for Musical Expression (NIME 2026), London, UK. Supplementary material included as an appendix. Code and demo: https://github.com/prabal-rje/latentscore Score: 5.30/10 (Obj:6 Id:5 Ind:3 Comp:5 Eff:5 Nov:5 Rep:8)

Strength: 完整开源的实时音乐接口系统，embedding retrieval backend 实现 ~1.2s CPU-only 延迟和 100% schema validity，live generator 架构有效隐藏 5-12s LLM 延迟，可复现性优秀（代码/数据/模型/benchmark 全部开源）。
Weakness: 核心评测指标 LAION-CLAP 同时参与 retrieval map 筛选和最终评测（循环论证），主定量结论（embedding lookup 优于 random，0.163 vs 0.139）缺乏独立验证，且无正式人类听感评估。
Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对现有基于GPU的文本转音频系统合成的是不可编辑的整段波形、难以实时交互控制这一问题，提出了一种实时音乐接口，能将自然语言场景描述转化为可演化的程序化声景。表演者输入诸如”午夜温暖爵士咖啡馆”的提示后，可通过直接参数调整来引导声音走向——降低明亮度、切换节奏风格——每次调整都产生可预期、可听辨的变化，无需重新输入提示。与合成完整波形的系统不同，该乐器生成人类可读的配置，使表演者能理解和操作声音生成的底层逻辑。该方法将语言模型的文本理解能力与程序化音频合成相结合，为实时音乐表演提供了一个可控、透明的文本驱动声音创作工具。

We present a real-time musical interface that converts natural-language scene descriptions into evolving procedural soundscapes. A performer types a prompt such as “warm jazz cafe at midnight” and steers it through direct parameter adjustments - stepping brightness down, switching a rhythm style - each producing a predictable, audible shift without re-prompting. Where GPU-bound text-to-audio systems synthesize monolithic waveforms, our instrument generates human-readable configurations over a ca

14. CNN Models for Microphone Array Covariance Matrix Upsampling and Acoustic Imaging

Authors: Marianthi Adamopoulou, Parthasaarathy Sudarsanam, David Diaz-Guerra, Meng Jiang, Archontis Politis et al.

Categories: eess.AS, cs.LG, cs.SD, eess.SP

| Published in the 2026 IEEE International Symposium on Artificial Intelligence for Instrumentation and Measurement (AI4IM), Amalfi, Italy, 2026 Score: 5.26/10 (Obj:6 Id:4 Ind:8 Comp:4 Eff:3 Nov:3 Rep:9)

Strength: 低 — 5 个架构是 channel 数 × FDC 位置的工程枚举，无 principled 设计；最优模型 Hybrid-FDC-Expanded 仅比 Base CNN 好 0.020 RMSE，且训练时长从 3h 飙到 10d，性价比未讨论。
Weakness: 绝对指标层面，RMSE 从 0.548 (random) 降到 0.432，相对提升 21%，但无可用性阈值论证——0.432 的 SCM 误差在声学成像里是什么水平？能否支撑下游 DOA/声源定位？论文未给。相对 baseline 层面，缺 DBPN 对比是 fatal：核心效用 claim “outperforms” 实际只 outperform 了 random guess。
Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对实际声学成像系统中传感器数量有限导致空间分辨率不足的问题，提出利用卷积神经网络（CNN）对麦克风阵列协方差矩阵进行上采样，从而在不增加硬件复杂度的前提下提升声学成像的空间分辨率。具体而言，作者将一个四面体4麦克风阵列虚拟上采样为球形32麦克风阵列，通过深度学习方法估计各通道间的协方差矩阵，并比较了五种神经网络结构在该任务上的表现。该方法将声学成像从依赖物理传感器阵列扩展为基于数据驱动的虚拟阵列重建，为低硬件成本下的高分辨率声场可视化提供了新范式。

Acoustic imaging visualization is a core methodology in acoustics, enabling spatial analysis of sound sources and acoustic scenes. However, limited sensor availability in practical systems motivate approaches that enhance spatial resolution without increasing the hardware complexity. In this paper, we focus on upsampling virtually a tetrahedral 4-microphone array to a spherical 32-microphone array by estimating the covariance matrices of the channels employing deep learning techniques. Five neur

15. AmbiDrop: Ambisonics-Based Array-Agnostic Neural Speech Enhancement

Authors: Michael Tatarjitzky, Vladimir Tourbabin, Boaz Rafaely

Categories: eess.AS

| Submitted to IEEE Transactions on Audio, Speech, and Language Processing Score: 5.16/10 (Obj:8 Id:5 Ind:8 Comp:5 Eff:5 Nov:4 Rep:5)

Strength: 在未见阵列上 AmbiDrop 维持 SI-SDR 4.77 dB 而 baseline 崩溃至 -15.08 dB，且模型从 1.2M 压缩到 11K 参数仅损失 2 dB，展示了强泛化和部署可行性。
Weakness: 核心方法已发表于会议论文 (ICASSP 2026)，本版仅扩展实验；且未与 DeFTAN-AA 等同类 array-agnostic 方法进行任何实验对比，无法定位方法的相对优势。
Full review: Claude Code 全文七公理审稿

中文摘要： 多通道深度神经网络显著提升了语音增强性能，但通常受限于固定的麦克风阵列几何结构，导致在未见或非规则配置上泛化能力较差。现有的阵列无关方法往往依赖高复杂度架构或大规模多样化数据集，仍难以泛化到分布外的阵列布局。本文提出AmbiDrop，一种基于Ambisonics空间音频表示的阵列无关神经语音增强方法，并对该方法进行了深入分析。AmbiDrop利用Ambisonics将不同阵列的信号统一编码为与几何无关的表示，从而降低模型对特定阵列结构的依赖，在跨阵列配置的泛化性能上展现出优势。

Multichannel Deep Neural Networks (DNNs) have significantly improved speech enhancement performance; however, they typically remain constrained by reliance on fixed microphone array geometries, leading to poor generalization on unseen or irregular configurations. Current array-agnostic approaches often rely on high-complexity architectures or massive, diverse datasets, yet they still struggle to generalize to out-of-distribution layouts. In this paper, we present an in-depth analysis of AmbiDrop

16. Speech Playground: An Interactive Tool for Speech Analysis and Comparison

Authors: Stephen McIntosh, Daisuke Saito, Nobuaki Minematsu

Categories: cs.CL, cs.SD, eess.AS

| Accepted to Interspeech 2026 (Show and Tell); 2 pages, 3 figures Score: 4.95/10 (Obj:5 Id:5 Ind:9 Comp:4 Eff:4 Nov:4 Rep:9)

Strength: 开源工具将多种语音表示（SSL、articulatory、phonological、discrete、variable-length）整合到统一的交互式 Web 界面中，支持 DTW 对齐和强制对齐，MIT 许可证，代码完全可复现。
Weakness: 2 页 Show and Tell 论文无任何用户评估、系统性能指标或与同类工具（Voxplorer、wav2scape等）的功能对比，所有功能模块均为已有技术的工程组合，未论证组件间 synergy 或设计选择的必要性。
Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对现有语音分析工具（如Praat）难以与现代深度学习表示集成及进行对比分析的问题，提出了一个名为Speech Playground的交互式语音可视化与对比工具。该工具采用Python后端与Web前端的架构组合，支持对连续特征、离散特征和变长特征等多种语音表示类型进行交互式探索。系统集成了TextGrid和强制对齐（forced alignment）等功能，便于研究人员在统一界面中分析和对齐语音数据。Speech Playground弥补了传统工具在深度学习表示兼容性和多特征对比方面的不足，为语音研究提供了更便捷的分析平台。

This paper presents Speech Playground, an interactive speech visualization and comparison tool. While existing tools such as Praat are excellent, it can be cumbersome to integrate them with modern deep learning representations and use them for comparison. Speech Playground addresses this by combining a Python backend with a web-based frontend for interactive exploration of multiple feature types, including continuous, discrete, and variable-length representations. It includes TextGrid and forced

17. Quantifying the Uncertainty of Blindly Estimated Room Embeddings Using a Dispersion-Calibrated Score

Authors: Yang Xiang, Philipp Götz, Emanuël A. P. Habets, Andreas Walther, Wenwu Wang et al.

Categories: cs.SD, cs.LG

| Accepted to INTERSPEECH 2026 Score: 4.66/10 (Obj:6 Id:5 Ind:3 Comp:5 Eff:5 Nov:5 Rep:5)

Strength: 主要是同组前作 MRL 的渐进工程改进，组件均有先例，无新机制解释。
Weakness: 存在明显的循环构造链。(i) Stage-3 的训练目标 δ（散度）由“干净锚点 vs 损坏视图”的余弦距离定义，而 Stage-2 已经用相同损坏族（SpecAugment/pink noise）作为隐式增广；(ii) 评测的“散度一致性 ρ(U, δ)”用的就是训练监督信号 δ 本身——这是 reward=evaluation 重叠。
Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对从混响语音中估计的房间嵌入不可靠的问题——即使说话人、房间及收发几何保持不变，语音内容与录制退化仍会改变嵌入表示，进而损害下游任务性能。作者提出一个无需下游任务监督的框架，可从混响语音中学习对语音内容变化鲁棒的房间嵌入，并同时给出表示级的不确定性评分。该嵌入以结构化的房间脉冲响应为锚点，使表示具备物理可解释性并对内容扰动保持稳定。所提不确定性分数经分散度校准，能够在盲估计条件下量化嵌入的可信程度，为下游应用的可靠性判断提供依据。

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 re