每日 arXiv

音频 · 语音 · 音乐 · 声学

每日新爬取的 arXiv 论文，按评分排序。中文摘要帮你快速判断是否值得深读。

2026-07-02
日期2026-07-02
已评分—
均分—
最高—

Daily Papers — 2026-07-02

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

1. Decomposer: Learning to Decompile Symbolic Music to Programs

Authors: Yewon Kim, Apurva Gandhi, David Chung, Graham Neubig, Chris Donahue

Categories: cs.LG, cs.AI, cs.SD

| Project page: https://yewon-kim.com/decomposer Score: 6.60/10 (Obj:8 Id:6 Ind:5 Comp:6 Eff:8 Nov:6 Rep:6)

  • Strength: 本文首次定义了符号音乐反编译任务，并通过两阶段 SFT+RL 框架，使 8B 开源模型在忠实度（Onset F1 提升 +0.16 到 +0.32）和可读性（Rubric 0.61-0.74 vs 启发式 0.05-0.09）上均显著超越 frontier LLMs，并在 4 个 held-out 数据集上保持泛化优势。
  • Weakness: 可读性 reward 与评估指标使用相同的 LLM judge 和相同的 12 项 rubric，构成训练-评测循环依赖；方法框架（SFT+execution-RL）直接改编自 SVG/CAD 反编译工作，方法层新颖性有限；缺少 GDPO vs GRPO 消融和独立人类可读性验证。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对从音乐演奏中恢复高层指令这一逆问题，提出Decomposer——一个用于符号音乐反编译的后训练框架，旨在从符号音乐中恢复出可执行、可编辑的音乐程序。具体实现上，作者将该任务实例化为MIDI到Strudel的反编译，模型以符号MIDI为输入，输出Strudel音乐编程语言编写的程序。该工作将音乐表演的反演问题转化为可执行代码生成任务，为音乐指令的编辑与重构提供了新的解决路径。

Musical performance involves executing a set of high-level musical instructions, yet recovering those instructions from the performance is a challenging inverse problem. We present Decomposer, a post-training framework for symbolic music decompilation: the task of recovering executable, editable music programs from symbolic music. We instantiate the task as MIDI-to-Strudel decompilation, where the model takes symbolic MIDI as input and produces a program in Strudel, a music programming language,

2. Unlocking Speech-Text Compositional Powers: Instruction-Following Speech Language Models without Instruction Tuning

Authors: Congrui Du, Yang Zhang, Kaizhi Qian, Shiyu Chang

Categories: cs.CL, eess.AS Score: 6.30/10 (Obj:8 Id:6 Ind:5 Comp:8 Eff:6 Nov:5 Rep:7)

  • Strength: 仅用 30k 小时语音预训练 + 一次权重加法即可构建指令遵循 SLM，在 QA 任务上接近 ASR+TextLLM 上限（OpenbookQA 86.59%），在 emphasis generation 上达 SOTA（F1 31.42），方法极简且展示了跨模型族（OLMo, LLaMA）泛化性。
  • Weakness: 核心方法直接继承 ChatVector/Task Arithmetic，新颖性增量有限；emotion understanding（52.70%）远落后于 FUN-AUDIO-CHAT（74.74%），关键设计主张”必须从 base 而非 inst 训练”缺乏实验消融，且推理依赖外部 ASR + format forcing 等工程模块。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 语音语言模型（SLM）的指令微调比文本大模型困难得多，既需学习新模态又需覆盖大量语音专属指令，而现有方法照搬文本范式合成大规模语音数据，因语音序列远长于文本而难以扩展。本文提出SpeechCombine，仅用一轮3万小时的语音预训练、完全不做指令微调，即可获得具备指令遵循能力的SLM：在文本基础模型上对语音做持续预训练得到语音适配模型，再将其权重与”指令微调版—基础版”文本LLM的权重差直接相加。实验表明，这种简单组合既保留了原文本LLM的知识与能力，又能将其有效迁移至语音领域。该工作为SLM训练指明了摆脱海量语音数据依赖的新方向。该论文针对语音语言模型（SLM）指令微调远比文本大模型困难这一问题——后者需同时学习新模态与大量语音专属指令——提出了一种无需指令微调即可解锁语音-文本组合能力的新方法。作者指出，现有SLM训练范式照搬文本LLM的做法，通过合成大规模语音预训练与指令微调数据集来训练，难以高效扩展。为此，本文绕开昂贵的指令微调流程，探索在不进行指令调优的前提下让SLM具备遵循指令的能力。该方法旨在释放语音与文本之间的组合互补潜力。实验表明，该路径能在显著降低数据与训练成本的同时实现具有竞争力的指令遵循表现。

如需更准确的概括，烦请提供完整英文摘要（您贴的内容在“speech sequence”处被截断了）。

Instruction tuning for speech language models (SLMs) is substantially more challenging than for text-based large language models (LLMs), as it requires learning a new modality and a wide range of speech-specific instructions in addition to those supported by text LLMs. Existing SLM training approaches largely replicate the text LLM training paradigm by synthesizing large-scale speech pre-training and instruction-tuning datasets. However, this strategy is difficult to scale, since speech sequence

3. SelectTSL: Prompt-Guided Selective Target Sound Localization in Complex Scenarios

Authors: Ziyang Jiang, Yu Chen, Zexu Pan, Xinyuan Qian, Bowen Xing et al.

Categories: cs.SD, cs.AI Score: 6.21/10 (Obj:8 Id:6 Ind:8 Comp:5 Eff:8 Nov:5 Rep:6)

  • Strength: 在双通道 prompt-guided selective SSL 任务上，SelectTSL 在合成数据上达到 MAE 0.98°/MOTA 91.57%，在 TAU-SRIR 真实数据上达到 MAE 2.62°/MOTA 0.77，全面超越 11 个 baselines，消融实验系统验证了各组件贡献。
  • Weakness: 架构复杂度高（PGSA+IPD Enhancer+PRTM+Cardinality Head，涉及 FiLM/cross-attention/DPRNN/TCN/BiGRU/SE 等大量标准模块堆叠），缺少与简单端到端设计的对比；未引用且未与最直接的 concurrent work LuSeeL (arXiv:2601.19153) 对比；代码和数据尚未开源。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 人类能在复杂场景中选择性关注目标声源并估计其方向，而现有深度学习系统难以实现这种选择性定位。现有声源定位(SSL)方法通常定位所有声源而不具备选择性，目标声源提取(TSE)虽能利用多模态提示提取目标声源但常丢失多通道空间信息。为此，本文提出SelectTSL方法，通过提示引导实现复杂场景下的选择性目标声源定位。该方法在保留多通道空间信息的同时，借助多模态提示选择性定位目标声源，弥补了SSL与TSE两类任务的不足。实验表明SelectTSL在复杂场景下的选择性目标声源定位任务上取得了有效成果。

Humans can selectively attend to a target sound and estimate its direction in complex scenarios, whereas such selective localization remains challenging for current deep learning-based systems. Sound source localization (SSL) has achieved remarkable success with deep learning, yet most methods localize all active sources without selectivity. Conversely, target sound extraction (TSE) extracts sources using multimodal prompts but typically fails to preserve the multichannel spatial information req

4. Self-Supervised Test-Time Tuning for Packet Loss Concealment

Authors: Yehoshua Dissen, Joseph Keshet

Categories: eess.AS, cs.CL

| Under submission to IEEE TASLP Score: 6.10/10 (Obj:8 Id:6 Ind:9 Comp:5 Eff:6 Nov:5 Rep:8)

  • Strength: 在两个开源 PLC backbone（FRN, PARCnet）上系统验证了合成掩码自监督测试时调优框架，含非因果/因果部署，FRN 因果 replay 恢复非因果增益的 77-102%，并用 bit-exact canary test 严格验证无数据泄漏。
  • Weakness: 绝对质量改进有限（PESQ +0.04, PLCMOS-NI +0.037 on LibriSpeech-40；PARCnet OOD 语音 PEAQ 仍在 -3.75 区间），缺少关键 ablation（burst depth, held-out ratio, fold count）和与更简自适应方法及更强 PLC SOTA（Diff-PLC, Flow-PLC）的对比。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 针对实时音频通信中的丢包隐藏问题，现有方法通常在部署时保持模型参数固定，忽略了每个通话中已接收数据包所携带的信号特定信息这一有用线索。为此，本文提出了TTT-PLC，一种自监督的测试时调优框架，能够仅利用接收到的数据包对已有的PLC模型进行在线自适应。该方法通过对已接收的音频片段进行合成掩码来生成自监督信号，从而在推理阶段动态更新模型参数。实验表明，TTT-PLC能够显著提升多种现有PLC模型的 concealment 质量，验证了测试时自适应在实时音频恢复中的有效性。

Packet loss concealment (PLC) reconstructs audio packets that are missing at the receiver, usually with a trained model whose parameters remain fixed at deployment time. This treats the PLC model as static, even though each call or recording exposes signal-specific information through the packets that did arrive. We present TTT-PLC, a self-supervised test-time tuning framework that adapts existing PLC models using only those received packets. The method creates supervision by synthetically maski

5. H-SAGE: Holistic Speaker-Aware Guided Experts for MoE-based Multi-Talker ASR

Authors: Yujie Guo, Jiaming Zhou, Yuhang Jia, Yang chen, Yong Qin

Categories: cs.SD Score: 6.00/10 (Obj:8 Id:6 Ind:8 Comp:5 Eff:6 Nov:5 Rep:8)

  • Strength: 在 GLAD 基础上引入显式 overlap 监督和 holistic gating，在 LibriSpeechMix 高重叠场景和 zero-shot 3-mix 泛化上取得一致性提升（WER 5.7 vs GLAD 6.2 on 2-mix, 19.5 vs 20.0 on 3-mix），代码开源。
  • Weakness: 三个核心组件（self-attention encoder、三态 overlap 辅助损失、concat+softmax 门控）均为已有技术的直接组合，baseline 仅对比 3 个系统（遗漏 SA-SOT/BA-SOT/SD-CTC/SSA 等近期工作），最简 baseline（能量阈值 overlap detector）缺失，3-mix 绝对 WER 19.5% 距实用水平仍有差距。
  • Full review: Claude Code 全文七公理审稿

中文摘要： H-SAGE 针对基于 MoE 的多说话人语音识别中帧级独立路由导致的时序短视和仅依赖 ASR 目标造成的表征学习不接地这一问题，提出了整体性说话人感知引导专家框架。该方法通过引入 Speaker-Aware Global Encoder 捕获长时依赖，并以 Overlap-Aware Loss 作为辅助监督显式引导模型区分声学状态；同时设计了 Holistic Gating Mechanism，联合评估全局上下文与局部细节来仲裁专家选择。在 LibriSpeechMix 数据集上的实验表明，H-SAGE 相较强基线取得了一致性的提升，尤其在复杂高重叠场景下表现突出，验证了显式声学引导能有效增强专家协作。> Multi-talker Automatic Speech Recognition (MTASR) faces significant challenges in accurately transcribing overlapping speech, particularly under complex high-overlap conditions. While recent Mixture-of-Experts (MoE) approaches have shown promise, they typically rely on frame-independent routing that leads to temporal myopia, and depend solely on the downstream ASR objective, which results in implicit and ungrounded representation learning. To address these limitations, we propose Holistic Speake

6. Neural Audio Codec with Adjustable Token Temporal Resolution Using Sampling-Frequency-Independent Convolutional Layers

Authors: Tomohiko Nakamura, Wataru Nakata, Kanami Imamura, Yuki Saito

Categories: eess.AS, cs.SD

| Accepted for IWAENC 2026 Score: 5.95/10 (Obj:8 Id:6 Ind:8 Comp:8 Eff:5 Nov:5 Rep:6)

  • Strength: 仅替换 DAC 中 quantizer 前后两个卷积层为 SFI 层（每 channel pair 仅 3 参数 vs Naive 的 231 参数），即可在 7 种 TTR 上全面优于 TTR-specific 层切换 baseline，并在大 TTR 时接近单独训练的 Reference 模型。
  • Weakness: 在小 TTR（13.3–26.7ms）下与 Reference 差距显著且未报告具体数值，仅在环境声数据集上验证（未覆盖语音/音乐），缺少与 SNAC、FlexiCodec 等同期多 TTR 方案的直接对比和下游任务验证。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 神经音频编解码器（NAC）生成的离散token在音频生成与理解模型中被广泛用作紧凑表示，其中token时间分辨率（TTR）决定了相邻token帧之间的时间间隔，需在精确刻画快速声学事件与缩短token序列长度之间取得平衡。然而现有大多数NAC仅在单一TTR下训练，为不同TTR需重复训练，成本高昂。本文提出一种采用采样频率无关卷积层的NAC架构，使模型能够在推理时灵活调整TTR而无需为每种分辨率单独训练。该方法将TTR与编解码器结构解耦，通过统一框架支持多种时间分辨率配置。实验表明，该编解码器在可调TTR设置下能在重建质量与token压缩率之间实现与固定TTR模型相当甚至更优的权衡。

Discrete tokens obtained from neural audio codecs (NACs) have been used as compact representations in audio generation and understanding models. In such token-based systems, token temporal resolution (TTR), defined as the time interval between adjacent token frames, is important because it controls the trade-off between representing rapid acoustic events and reducing token-sequence length. However, most NACs are trained at a single TTR and require separate training for each TTR. This paper propo

7. Speaker head orientation estimation with a single microphone array using phase spectrogram features

Authors: Balint Turi, Archontis Politis, Parthasaarathy Sudarsanam, Tuomas Virtanen

Categories: cs.SD

| Accepted to EUSIPCO 2026 Score: 5.90/10 (Obj:8 Id:5 Ind:6 Comp:5 Eff:8 Nov:5 Rep:6)

  • Strength: STFT 相位特征在模拟和真实数据上均大幅优于 baseline（模拟 clean 19.9° vs 44.8°，真实数据 73.2% vs 65.4%），且个性化微调达到 11.3° MAE。
  • Weakness: 核心特征表示（STFT phase → CNN）和模型架构（CRNN+Attention）均直接迁移自 SELD/DOA 领域已有工作，且未引用高度相关的 arXiv:1705.00919，消融实验不足以识别各组件贡献。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对智能环境、会议和驾驶员监控等场景中从音频估计说话人头部朝向的问题，提出了一种仅依赖单一麦克风阵列的新方法。该方法以短时傅里叶变换的相位谱图作为输入，结合卷积、循环和自注意力层构建深度神经网络进行学习，区别于以往使用物理启发式手工特征或原始波形输入的方案。这种基于相位特征端到端学习的方式使模型能够从数据中鲁棒地建模头部朝向与声场之间的关系，避免了手工特征设计的局限性。该工作为无需视觉传感器的说话人朝向估计提供了新的解决思路。

Estimating a speaker’s head orientation from audio can provide valuable information in smart environments, meetings, and driver monitoring. We propose a novel approach that leverages the phase component of the short-time Fourier transform from a single microphone array as input to a deep neural network combining convolutional, recurrent, and self-attention layers. Unlike prior methods that use physics-informed handcrafted features or raw waveform inputs, our approach enables robust learning from

8. Rethinking Speech-LLM Integration for ASR: Effective Joint Speech-Text Training by Interleaving

Authors: Ruchao Fan, Yiming Wang, Rui Zhao, Liliang Ren, Keqi Deng et al.

Categories: cs.CL, eess.AS Score: 5.79/10 (Obj:8 Id:6 Ind:5 Comp:8 Eff:6 Nov:5 Rep:3)

  • Strength: 中 — word-level concat-then-reinsert 工程方案有实用价值，Mixed-IL 经验有效但机制解释不足。
  • Weakness: 几乎不可独立复现：(1) 主训练数据 38k hours in-house ASR、TTS-pairs、medical/banking test sets、entity 标注均不公开；(2) backbone 是 internal 7B LLM（5T tokens 训练），非公开模型；(3) 评测脚本、EER 标注规范、speech-MMLU 构造细节未公开；(4) 论文未提及代码/checkpoint 发布计划；(5) HMM ali。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 语音与大语言模型(LLM)的集成虽能借助丰富的文本预训练获得收益，但其对自动语音识别(ASR)的具体贡献仍不清晰。作者观察到，随着监督ASR训练数据量的增加，LLM先验的作用逐渐减弱，而简单的语音-文本联合训练又未能充分利用文本知识。为此，该论文提出了一种面向ASR的联合语音-文本交错预训练策略JSTIP，通过在词级等粒度上构建语音与文本的交错序列进行联合训练。该方法旨在更有效地将LLM的文本知识迁移到语音识别任务中，使文本先验在数据规模扩大时仍能持续发挥贡献，从而提升ASR的整体识别性能。

Speech-LLM integration has shown promising results by leveraging extensive textual pretraining, yet its specific benefits for automatic speech recognition (ASR) remain unclear. We observe that as supervised ASR training data increases, the contribution of LLM priors becomes less evident, and simple speech-text joint training under-utilizes textual knowledge. We therefore propose Joint Speech-Text Interleaved Pretraining (JSTIP), an ASR-oriented pretraining strategy that constructs word-level and

9. Audio-Based Understanding of Audiobook Narration Appeal

Authors: Shahar Elisha, Mariano Beguerisse-Díaz, Emmanouil Benetos

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

| Accepted to Interspeech 2026 Score: 5.74/10 (Obj:6 Id:5 Ind:8 Comp:6 Eff:5 Nov:6 Rep:5)

  • Strength: 低 — 全部为 standard tool 拼装，无新方法、无新表示、无新评测协议；贡献是经验发现而非方法增量。
  • Weakness: 论文使用 GLM + LME（book-group 随机截距）隔离 title 效应，per-genre GLM 隔离 genre 效应，VIF 剪枝处理多重共线性，BH 校正多重比较。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 本文研究有声书叙述质量的吸引力问题，探讨叙述特征如何影响听众的收听体验与内容理解。作者从 LibriVox 数据集中利用预训练音频模型提取声调和语速等声音与声学特征，并分析这些特征与消费数据（特别是收听率）之间的关系及其与体裁和书名的交互作用。研究发现叙述质量的影响因体裁、书名和受众不同而存在差异，揭示了有声书叙述吸引力的多维度特征。该工作为理解有声书听众行为和优化叙述策略提供了数据驱动的分析视角。

Narration is central to the audiobook listening experience, shaping how listeners engage with and understand the content. This work explores how narration qualities shape an audiobook’s appeal, noting that their effects can vary by genre, title, and audience. We extract vocal and acoustic features (e.g., tone, pace, loudness) from LibriVox using pre-trained audio models and analyse their relationship with consumption data (specifically, view-rate) and their interplay with genre and title. Despit

10. Reasoning LLM Improves Speaker Recognition in Long-form TV Dramas

Authors: Yuxuan Li, Lingxi Xie, Xinyue Huo, Jihao Qiu, Jiacheng Shao et al.

Categories: cs.CL, cs.AI, cs.CV

| Accepted to ICML 2026 Score: 5.60/10 (Obj:8 Id:5 Ind:5 Comp:5 Eff:5 Nov:5 Rep:8)

  • Strength: 构建了首个大规模长篇电视剧说话人识别 benchmark（DramaSR-532K, 532K utterances, 900+ 角色, 525h 视频），人工标注质量高（99.6% inter-annotator agreement, 0.5% label noise），并通过 reasoning LRM + 多模态工具调用在 hard cases（short utterances +9.20%）上获得显著提升。
  • Weakness: 核心方法提升仅 +2.30% absolute（85.49%→87.79%），且 baseline（Label Propagation）为自建而非独立 SOTA；缺少与 SpeakerLM (AAAI 2026)、DiarizationLM 等直接竞争方法的性能比较；pipeline 涉及 4+ 模型但缺少模块必要性的充分消融。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 长篇电视剧的视频理解中，将每句台词准确归属到对应角色的”说话人识别”是理解复杂剧情的关键难题。本文提出两项核心贡献：一是构建了大规模基准数据集 DramaSR-532K，涵盖 532K 条标注对白及 900 余个独特角色，要求模型综合多模态线索进行人物识别；二是引入具备推理能力的大语言模型（Reasoning LLM）来提升说话人识别的准确率。该工作通过将推理能力与多模态信息融合，在长剧情场景中实现了更稳健的角色-台词归属。DramaSR-532K 为该方向的首个大规模系统化评测基准，填补了长视频理解中说话人识别任务的空白。

Long-form TV dramas present a formidable challenge for comprehensive video understanding, where deciphering complex storyline often relies on \textbf{speaker recognition}, the task of accurately attributing each spoken utterance to its respective character. In this paper, we advance this field through two primary contributions. (1) We introduce \textbf{DramaSR-532K}, a large-scale benchmark comprising 532K annotated dialogue lines across more than 900 unique characters, necessitating the integra

11. Towards a Phonology-Informed Evaluation of Multilingual TTS

Authors: Sneha Ray Barman, Neeraj Kumar Sharma, Shakuntala Mahanta

Categories: cs.CL, cs.ET, cs.LG

| Accepted at Interspeech 2026 Score: 5.60/10 (Obj:6 Id:5 Ind:9 Comp:5 Eff:5 Nov:5 Rep:9)

  • Strength: 提出基于分类器的 TTS 音系忠实度审计框架，在阿萨姆语 ATR 元音和谐上发现 MMS TTS 的 [+ATR] 中元音有 ~33% 被误判为 [-ATR]，7:1 欠产出偏差在人类语音中不存在，代码和数据完全开源。
  • Weakness: 仅评估单一 TTS 系统、单一语言、单一音系现象，TTS 数据集仅 281 tokens 且类别严重不平衡，缺少与已有评估方法的直接对比和第二语言/现象的泛化验证。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 神经TTS系统虽能跨语言生成自然语音，但自然性并不等同于保留了区分词义和语法形式的语音对立，而MOS等主流评测指标对此无能为力。为此，论文提出一个基于分类器的评测框架，以人类语音为基准，依据语言特异的音系模式对TTS输出进行审计。作者以阿萨姆语的前舌根（ATR）元音和谐为测试场景，对Meta的MMS TTS展开实验，用以验证该框架能否捕捉传统指标遗漏的音系信息。结果表明，在人类语音上训练的分类器能够有效检验TTS输出是否忠实保留了特定音系对比，从而弥补了自然度评测的盲区。该工作为多语言TTS的音系学层面评测提供了一种可推广的诊断工具。

（注：因英文摘要尾部被截断，以上对主要结果的概括基于可见内容推断，具体数值结论请以原文为准。）

Neural TTS systems can sound natural across languages, but naturalness does not guarantee the preservation of sound contrasts that distinguish words from their grammatical forms. Standard metrics like MOS do not test for this. We propose a classifier-based framework that audits TTS output against language-specific phonological patterns using human speech as a benchmark. Testing Assamese advanced tongue root (ATR) vowel harmony with Meta’s MMS TTS, we show that a classifier trained on human speec

12. Enhancing Acoustic-to-Articulatory Inversion with Multi-Target Pretraining for Low-Resource Settings

Authors: Jesuraj Bandekar, Prasanta Kumar Ghosh

Categories: eess.AS Score: 5.60/10 (Obj:8 Id:5 Ind:9 Comp:5 Eff:5 Nov:4 Rep:9)

  • Strength: 在低资源 AAI 场景（6.25% 数据）下，多目标预训练使 seen speakers CC 从 0.7348 提升至 0.7811（+6.3%），unseen speakers CC 从 0.6687 提升至 0.7259（+8.5%），同时推理时无需 SSL 特征提取器。
  • Weakness: 三目标联合的必要性未被证明（7 种配置差异 <1%，且 ACP-T 非全局最优），仅对比 TERA 一种 SSL baseline（未覆盖 wav2vec2/HuBERT），高资源场景提升微弱（100% 数据 CC +0.2%），且未隔离预训练数据量对提升的贡献。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 本研究针对低资源场景下声学-发音反演（AAI）任务中自监督学习（SSL）特征提取器引入推理延迟与计算开销的问题，提出一种多目标预训练方法。该方法利用音素标签、发音特征标签和关键发音器官标签三种目标表示进行预训练，从而在推理阶段无需依赖SSL特征提取器。实验在多种数据条件下与基线模型及基于SSL的模型进行了对比评估。结果表明，该方法在低资源场景中持续提升AAI性能，同时在不牺牲精度的前提下显著降低了推理开销。> Acoustic-to-Articulatory Inversion (AAI) estimates vocal tract articulator movements from speech, benefiting tasks like ASR, speech synthesis, and speaker verification. While deep learning-based methods (CNNs, RNNs, Transformers) have advanced AAI, recent studies show that Self-Supervised Learning (SSL) features further enhance performance, particularly in low-resource settings. However, SSL feature extractors introduce inference latency and computational overhead. To address this, we propose a

13. A Multi-Branch Hierarchy-Aware Framework for Heterogeneous Audio Classification

Authors: Beile Ning, Jiayi Yu, Zitong Wang, Yufei Hu, Wenjun Xu et al.

Categories: cs.SD, cs.AI Score: 5.53/10 (Obj:9 Id:5 Ind:6 Comp:4 Eff:8 Nov:2 Rep:6)

  • Strength: 中 — 工程集成有效，逐组件 ablation 单调提升，log-STFT 主导贡献的发现对后续参赛者有参考价值；但无 synergy 证明，复杂度换取的增益中有 ~60% 来自数据扩充。
  • Weakness: 创新弱，主要是已知组件的组合。(1) CLAP 音频-文本表示 — Wu et al. ICASSP 2023，已用。(2) 层级分类头 (Flat/GC/LCL) — Ding et al. 2023 已在声学场景做 hierarchical classification，论文引用 [4] 并承认”inspired from [4]”；HiClass 库 (JMLR 2023) 显示 LCL 是标准层级分类技术。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对DCASE 2026挑战赛Task 1中异构音频分类问题，即在宽声音分类法（BST）下对音频进行二级准确预测并保持与顶层分类体系的一致性。作者基于CLAP音频-文本表征构建系统，并从三个方向加以改进：通过筛选BSD35k子集扩充训练数据、引入针对不同声学特征的特定分支增强建模能力、以及利用层次感知分类器和基于KNN的后处理细化预测结果。在所考察的声学特征中，log-STFT分支作为单模型表现最佳，配合KNN后处理在BSD10k-v1.2集上取得80.84%的层次F1分数。进一步通过融合具有互补声学特征与分类头的模型，集成系统分别达到81.25%和81.18%的层次F1分数。> This technical report describes our system for Task 1 of the DCASE 2026 Challenge, which aims to classify heterogeneous audio recordings according to the Broad Sound Taxonomy (BST). The task requires both accurate second-level prediction and consistency with the top-level taxonomy. Our system is built on CLAP-based audio-text representations and is improved along three strategies: expanding the training set with a filtered subset of BSD35k, enhancing acoustic modeling with feature-specific branc

14. Cross Domain Few-Shot Class-Incremental Audio Classification Via Adversarial Contrastive Learning

Authors: Yongjie Si, Yanxiong Li, Sen Huang, Beibei Liu

Categories: eess.AS

| 5 pages, 3 figures, 4 tables, accepted for publication in Interspeech 2026 Score: 5.50/10 (Obj:6 Id:5 Ind:8 Comp:5 Eff:6 Nov:4 Rep:8)

  • Strength: 在6对跨域音频数据集上AA均超越9个基线方法（提升+0.47%至+3.63%），消融实验确认对抗训练与监督对比损失有独立贡献且代码已开源。
  • Weakness: 核心方法是对抗训练+对比学习的已有组合（ASCL 2022, CDA 2023等先期工作）的直接迁移，跨域FSCIL问题设置在医学影像和点云领域已有先例，且在FS→NS和FS→LS方向绝对精度仅41-47%远低于可用水平。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 现有少样本类增量音频分类（FCAC）方法假设基类与增量类样本服从同一分布，但在实际场景中两类样本之间往往存在域偏移问题。本文针对这一不足，研究了跨域少样本类增量音频分类（Cross Domain FCAC）问题，即基类与增量类样本之间存在分布差异的设定。为此，作者提出了一种对抗对比学习策略，通过对抗训练与对比学习的结合来缓解域偏移带来的负面影响，使模型在增量学习新类别时仍能保持对基类的判别能力。该方法为跨域增量音频分类提供了新的解决思路，有助于提升模型在域差异条件下的分类性能与泛化能力。

Current Few-shot Class-incremental Audio Classification (FCAC) methods assume that samples of base and incremental classes are in the same domain (following the same distribution). However, there is generally a domain shift between the above two types of samples. In this paper, we explore the problem of Cross Domain FCAC where samples of base and incremental classes have domain shift. We propose a strategy of adversarial contrastive training which enables the model to effectively classify sample

15. Pmeta-TLA: Backdoor Attacks for Speech Classification Models via Meta-Learning with Timbre Leakage Attack

Authors: Yueming Huang, Wenhan Yao, Fen Xiao, Xiarun Chen, Weiping Wen

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

  • Strength: TLA 触发器利用音色帧级嵌入自监督特征空间，在 10 个 baseline 上取得 95%+ ASR 且被 DNN 检测器识别率降低 30-50%；Pmeta-TLA 的多后门注入在 5 后门场景下保持 >90% ASR。
  • Weakness: 缺少 LFBA (IEEE SMC 2025) 和 LRBA (Interspeech 2025) 两个最接近的帧级/潜空间攻击 baseline；消融实验未隔离”音色信息”与”帧级扰动”两个变量，且未提供 TLA 与 Pmeta-TLA 之间的 synergy 证明；代码未开源。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对语音分类模型在后门攻击中的安全问题，指出现有语音触发器容易被深度神经网络检测器识别的脆弱性，提出了一种新的音色泄露攻击（Timbre Leakage Attack, TLA）方法。该方法利用语音信号中的音色特征作为后门触发器，使得攻击更具隐蔽性，难以被现有防御机制发现。在此基础上，作者进一步结合元学习策略（Pmeta-TLA），优化触发器的生成与植入过程，提高攻击在不同模型和数据集上的泛化能力与稳定性。实验结果表明，所提攻击方法在保持高攻击成功率的同时，显著降低了被检测器识别的风险，揭示了当前语音分类模型在安全防护上的潜在缺陷，并为后续防御研究提供了重要的参考基准。

Recently, speech classification methods have gained widespread adoption in intelligent gadgets. Current study indicates that backdoor attacks provide a substantial security concern to these models, underscoring the pressing necessity to investigate additional potential attack techniques to expose and prevent such risks. This work discusses the vulnerability of current speech triggers to detection by deep neural network defenders and introduces the Timbre Leakage Attack (TLA). The suggested trigg

16. Spatial Speech Perception Systems: A Survey of Sound Source Localization, Directional Enhancement, and Speech Recognition

Authors: Pengyuan Shao, Dimitrios Kanoulas

Categories: eess.AS

| 27 pages, 2 figures, 7 tables. Survey paper Score: 5.30/10 (Obj:7 Id:5 Ind:5 Comp:5 Eff:6 Nov:5 Rep:4)

  • Strength: 首次将 SSL、DSE、ASR 三大子领域纳入统一 pipeline 视角综述，Table V-VII 汇编了 10 个 SSL 系统的实时性和噪声鲁棒性数据及 8 个 ASR 模型的延迟数据，有实用参考价值。
  • Weakness: 9 页 / 106 篇引用对三领域综述严重不足，无系统化文献筛选方法（无 PRISMA 协议、无搜索策略描述），分析为描述性目录而非因果综合，且遗漏 Huang et al. 2025 等关键同领域综述。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 本文综述了空间语音感知系统的研究进展，聚焦于在真实复杂声学环境（如机器人听觉、助听器、电话会议、智能音箱等场景）下实现稳健语音理解的核心难题，即背景噪声、混响、竞争说话人及动态声学条件对系统性能的影响。文章系统梳理了三大关键技术：基于麦克风阵列的声源定位、目标语音的方向性增强，以及空间信息辅助的语音识别，构建了从感知到理解的完整技术链条。作者回顾了传统信号处理方法与近年来基于深度学习的空间语音处理方法的演进，并讨论了多通道空间特征建模、跨模态融合及端到端优化等前沿方向。综述还总结了公开数据集、评估指标与基准对比，为研究者提供了可复现的实验参考。最后，文章指出了真实场景泛化、低延迟实时处理及多说话人复杂交互等尚未解决的开放问题，并展望了空间语音感知系统未来在具身智能与人机交互中的应用前景。

Robust speech understanding in real-world acoustic environments remains a fundamental challenge for intelligent auditory systems such as robot audition, hearing aids, teleconferencing systems, smart speakers, and voice-controlled assistants. These systems must operate under background noise, reverberation, competing speakers, and dynamic acoustic conditions. Spatial speech perception addresses this challenge by exploiting microphone-array information to localize, enhance, and interpret target sp

17. DRL-CLBA: A Clean Label Backdoor Attack for Speech Classification via DDPG Reinforcement Learning

Authors: Yueming Huang, Wenhan Yao, Fen Xiao, Xiarun Chen, Weiping Wen

Categories: cs.AI, cs.SD Score: 5.30/10 (Obj:6 Id:5 Ind:6 Comp:3 Eff:6 Nov:5 Rep:5)

  • Strength: 在 7 个语音数据集 × 4 种模型 × 3 类任务上，DRL-CLBA 的 ASR 持续超越 5 个 baseline（KWS 平均 88-90%，SV 88%，SER 77-79%），BA 基本无损，且对 fine-tuning 和 pruning 有一定抵抗力。
  • Weakness: 摘要声称抵抗 spectral signature 防御但实验未评估；核心创新为 CSSBA 框架 + DDPG 优化器的组件替换，但缺少 DDPG vs 调优良好的同步数 PGD 的关键隔离对照，无法判定增益来自 DDPG 序列建模还是超参/网络容量差异。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对语音分类模型中后门攻击的标签污染问题，提出了一种基于深度确定性策略梯度（DDPG）强化学习的干净标签后门攻击方法DRL-CLBA。该方法通过深度音频特征扰动生成样本特定的触发器，在不修改目标样本标签的情况下注入后门，从而规避人工数据审查防御。实验表明，DRL-CLBA在保持高攻击成功率的同时，显著提升了触发器的隐蔽性和对现有防御手段的抵抗能力，验证了干净标签攻击在语音领域的有效性。

Deep learning models for speech classification are vulnerable to backdoor attacks, where malicious triggers cause misclassification at inference time. While sample-specific attacks can bypass many defenses, they often rely on poisoned label attack, making them detectable via manual data defense. In this paper, we propose DRL-CLBA, a novel clean label backdoor attack for speech classification that leverages Deep Deterministic Policy Gradient (DDPG) reinforcement learning. We also utilize deep aud

18. RT-Tango: Real-Time Distributed Binaural Speech Enhancement for Low-Power Hearing Aid Devices

Authors: Z. Benslimane, P. Chouteau, M. Poreba, F. Auzanneau, M. Szczepanski et al.

Categories: cs.SD

| Accepted at Interspeech 2026 Score: 5.20/10 (Obj:8 Id:5 Ind:8 Comp:5 Eff:5 Nov:4 Rep:5)

  • Strength: 在分布式双耳助听设备约束下将计算复杂度降至 33.41 MMACs/s（比 Tango-RNN 低 2x、比 GTCRN 低 6x）同时保持可比的 PESQ/STOI，并实现 8 ms 算法延迟的严格因果流式推理。
  • Weakness: 方法是四个已有组件（ERB 压缩、分组 RNN、固定跳帧、非对称 STFT）的工程组合而无新机制或新经验规律，且遗漏 Westhausen et al. (2024, TASLP) 和 LBCCN (Wang et al., ICASSP 2025) 两个直接竞争 baseline。
  • Full review: Claude Code 全文七公理审稿

中文摘要： RT-Tango针对资源受限助听设备上实时双耳语音增强的延迟、算力与设备间通信瓶颈，提出了一种两阶段分布式流式架构。该框架结合基于ERB的感知特征压缩，在低功耗平台上实现双耳协同处理。方法兼顾实时性与双耳空间线索的保持，在保证低延迟的前提下完成流式语音增强。实验表明该系统在计算开销与增强效果之间取得了有效平衡。

Real-time binaural speech enhancement is constrained by latency, computational cost, and inter-device communication, yet existing efficient solutions predominantly address single-channel settings. In this paper, we introduce RT-Tango, a real-time distributed binaural speech enhancement framework designed for streaming on resource-constrained platforms and specifically for hearing aids. RT-Tango relies on a two-stage distributed architecture combining perceptually motivated ERB feature compressio

19. LMPAN: A Lightweight Multi-Path Alignment Network for Joint Full-Duplex Acoustic Echo Cancellation and Noise Suppression

Authors: Chengwei Liu, Shaofei Xue, Haoyin Yan, Xiaotao Liang, Zheng Xue

Categories: eess.AS

| Accepted by Interspeech 2026 Score: 5.05/10 (Obj:6 Id:5 Ind:5 Comp:5 Eff:7 Nov:4 Rep:5)

  • Strength: 中 — 在 480K/126M MACs 预算下达到 SOTA-competitive,效率-质量 Pareto 改进真实,对工业部署有参考价值;但”comparable to DeepVQE-S”无证据,baseline 漏 CAGCRN/EchoFree。
  • Weakness: 核心 idea “多路对齐 + attention fusion + SSL 两阶段 + dynamic target” 是 A/B/C/D 组合,每个组件都有直接先例:soft time alignment → SCA [16], Align-ULCNet [15];attention fusion of LAEC+mic → DeepVQE [8] 内部已用;WavLM SSL 两阶段 → EchoFree 23;dynamic t。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对全双工语音对话系统中设备端联合声学回声消除（AEC）与噪声抑制（NS）所面临的硬件失真和动态声学条件问题，提出了一种轻量级多路径对齐网络（LMPAN）。该方法引入三个核心创新：多路径对齐模块用于校正参考信号、线性AEC输出与麦克风信号之间的时间与能量失配，基于注意力机制动态融合增强特征，并采用轻量化结构设计以满足端侧部署需求。实验表明，LMPAN在复杂声学环境下能有效提升联合回声消除与降噪性能，同时保持较低的计算开销与参数量。该工作为实时全双工对话场景提供了一种兼顾效果与效率的实用解决方案。

We propose a lightweight multi-path alignment network (LMPAN) for on-device joint acoustic echo cancellation (AEC) and noise suppression (NS) in full-duplex spoken dialogue systems. To address hardware-induced distortions and dynamic acoustic conditions, we introduce three core innovations: (1) a multi-path alignment stage correcting temporal and energy mismatches across reference, linear AEC (LAEC) output, and microphone signals; (2) an attention-based mechanism that dynamically integrates enha

20. UT-AISTimprt submission for ICME 2026 Grand Challenge on Academic Text-to-Music Generation

Authors: Shunsuke Yoshida, Yu-Hua Chen, Satoru Fukayama

Categories: cs.SD

| Accepted to ICME 2026 Grand Challenge on Academic Text-to-Music Generation Score: 4.95/10 (Obj:7 Id:4 Ind:7 Comp:5 Eff:5 Nov:3 Rep:5)

  • Strength: 低 — 直接迁移 CommonIT，增量仅为模态选择和 K 值扫描，无新方法贡献。
  • Weakness: 核心 idea（cluster samples by embedding similarity, build mini-batch from single cluster to reduce gradient interference）直接来自 CommonIT [7]，论文 Sec I 明确说 “Inspired by CommonIT, this work investigates clustering-based batch sa。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对低数据量与小规模模型设定下的文本到音乐生成任务，研究了训练阶段批采样策略的影响，作为ICME 2026学术文本到音乐生成大挑战赛的参赛方案。核心方法是对训练数据分别使用文本嵌入或音频嵌入进行聚类，将特征相似的样本分配到同一迷你批次中，从而缓解不同样本间梯度更新相互干扰的问题。通过对比不同模态嵌入与聚类策略对生成质量的作用，作者探究了在资源受限条件下如何通过批次构造方式提升模型训练效率与音乐生成表现。该工作为小规模文本到音乐生成系统在数据稀缺场景下的训练优化提供了实践参考。

This work investigates the effect of batch sampling strategies during training for text-to-audio music generation under low-data and small-scale model settings. This paper describes our approach and findings for the ICME 2026 Grand Challenge on Academic Text-to-Music Generation. Training data are clustered using either text embeddings or audio embeddings, and samples with similar characteristics are grouped within the same mini-batch to mitigate gradient interference. The effects of modality and

21. An Efficient vLLM-Based Inference Pipeline for Unified Audio Understanding and Generation

Authors: Haoran Wang, Jinchuan Tian, Siddhant Arora, Shinji Watanabe

Categories: eess.AS, cs.AI Score: 4.80/10 (Obj:8 Id:2 Ind:6 Comp:5 Eff:5 Nov:4 Rep:8)

  • Strength: 提出基于 vLLM 的 SpeechLM 推理管线，通过 Primary-Auxiliary Decomposition 和 Paired Request Co-Scheduling 在单 H100 上实现 ~108× 吞吐量提升并维持 CFG 80% 吞吐量，代码开源。
  • Weakness: 唯一 baseline 为顺序 PyTorch 推理，遗漏 VoxServe 和 vLLM-Omni 等直接竞品，生成质量存在退化（TTS WER 从 2.7→3.3），且 CFG 对比的 baseline concurrency 设置可能不公平。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 当前主流的高吞吐推理引擎（如vLLM）缺乏对多模态生成的原生支持，这一问题在语音语言模型中尤为突出：通过解耦的AR+NAR或同步多Token预测（MTP）配合延迟模式交错来生成多层音频Token，与标准单流解码循环存在结构性冲突。针对这一矛盾，作者提出了一种基于vLLM的统一语音理解与生成推理管线，将自回归解码扩展为可原生处理多层音频Token生成的形式，从而在同一框架内同时支持理解与生成两类任务。该方法克服了现有推理引擎在多模态生成场景下与单流循环不兼容的瓶颈，使语音语言模型能够在通用高吞吐推理框架中高效完成端到端的音频理解与生成，为多模态大模型在语音领域的部署提供了统一且高效的解决方案。

While Large Multimodal Models excel in comprehension, high-throughput inference engines lack native support for multimodal generation. This is severe in Speech Language Models, where generating multi-layered audio tokens via decoupled AR+NAR or synchronous Multi-Token Prediction (MTP) with delay-pattern interleaving conflicts with standard single-stream loops. We present a vLLM-based inference pipeline for unified speech understanding and generation. We extend autoregressive decoding to natively

22. Using embeddings to predict spoken word duration and pitch in Mandarin monosyllabic words

Authors: Xiaoyun Jin, Mirjam Ernestus, R. Harald Baayen

Categories: cs.CL Score: 4.74/10 (Obj:5 Id:4 Ind:8 Comp:5 Eff:5 Nov:3 Rep:5)

  • Strength: 低 — 方法完全沿用本组既有 DLM 线性映射，无新机制、无新压缩、无新 baseline。
  • Weakness: The paper explicitly positions itself as extending [10] (Gahl & Baayen 2024, English homophone duration from embeddings) and [12, 13] (Chuang/Lu, Mandarin f0 from CEs). The method (DLM linear map on CEs) is identical to。
  • Full review: Claude Code 全文七公理审稿

中文摘要： 该论文探究上下文嵌入（CEs）能否预测汉语口语单音节CV词的时长和音高（f0）。研究者从汉语 spontaneous 语料库中提取7470个单音节CV词token，分析其时间归一化f0曲线和时长特征。结果表明，CEs对时长的预测显著高于随机水平，不仅适用于词型层面，也适用于个体token层面，证明上下文嵌入编码了汉语口语中词汇的时长和音高信息。

Time-normalized f0 contours of Mandarin words in conversational speech have been shown to be predictable in part from their contextualized embeddings (CEs). The present study investigates whether CEs also predict spoken word duration for 7470 tokens of Mandarin monosyllabic CV words extracted from a Mandarin corpus of spontaneous speech. We show that CEs indeed are predictive for duration, above chance level, not only at the type level, but also at the level of individual tokens, as indicated by