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。

2026-07-13
日期2026-07-13
已评分—
均分—
最高—

Daily Papers — 2026-07-13

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

1. Encoder-Side Neuron Identification and Amplification for Acoustic Perception in Large Audio-Language Models

Authors: Yu-Han Huang, Chih-Kai Yang, Ke-Han Lu, An-Yu Cheng, Hung-yi Lee

Categories: cs.SD, cs.AI Score: 8.05/10 (Obj:9 Id:9 Ind:8 Comp:8 Eff:9 Nov:7 Rep:6)

  • Strength: 首次在音频编码器内部进行神经元级免训练干预（IAAN），仅放大 ≤0.12% 编码器神经元即在三个开源 LALM 上将非语义语音属性准确率提升 9.7–25.7 个百分点，同时 ALA 下降 21–35 点；严格控制实验（Random/Layer-amplify/LLM-amplify/AAD/HS-steer）同时隔离了干预位置和神经元选择性两个因果变量。
  • Weakness: 未引用两篇高度相关先行工作（arXiv:2601.03115 对同一组模型做标签依赖神经元级情绪干预、arXiv:2602.15307 对音频 SSL 模型做神经元级分析），且论文未提供代码开源链接，噪声参考生成的完整伪代码缺失。
  • Full review: Claude Code 全文七公理审稿

Large audio-language models (LALMs) often underperform on fine-grained, non-semantic attributes of speech, such as a speaker’s emotion, despite strong performance on speech content. Improving this without the cost of retraining calls for an effective inference-time intervention, yet most existing methods intervene only after the audio encoder and operate at a relatively coarse granularity. The encoder itself, where acoustic information is first extracted from the waveform, remains largely unexplored, especially at the level of individual neurons. We introduce IAAN, Identifying and Amplifying Acoustic Neurons, a training-free and label-free method that scores each feed-forward neuron in the audio encoder by contrasting its activation on the real waveform with that on a noise reference lacking the real audio’s acoustic information. IAAN then amplifies a small set of the highest-scoring neurons at inference. Across ten non-semantic speech attributes, IAAN improves average accuracy by 25.7 points on Audio-Flamingo-3, 21.4 on Qwen2.5-Omni, and 9.7 on Kimi-Audio. It also improves a model already explicitly fine-tuned to prioritize acoustic evidence. In controlled comparisons, both the encoder locus and neuron-level selectivity prove necessary for this gain. Intervening after the encoder, at the decoding side or inside the language model, yields little to no improvement, or even deteriorates accuracy. The improvement also depends on which specific neurons are amplified, not merely on their number, confirming that IAAN’s acoustic score succeeds in identifying the neurons that matter. These results show that a small, precisely targeted intervention inside the audio encoder is an effective and largely untapped way to strengthen the acoustic understanding of LALMs, opening a new direction for inference-time methods that improve acoustic perception through neuron-level access to the encoder.

2. BeatEdit: Symbolic Music Generation as Explicit Editing

Authors: Haoyu Gu, Lekai Qian, Haowu Zhou, Qi Liu, Shuai Wang

Categories: cs.SD, cs.AI Score: 8.00/10 (Obj:9 Id:8 Ind:6 Comp:8 Eff:9 Nov:7 Rep:9)

  • Strength: 首个基于显式编辑操作的符号音乐生成框架，通过 R1-R3 表征分析和 2×2 编码-方法析因实验揭示编码选择为重要设计杠杆，纠错 beat exact match 0.726 近 2× 于最强生成 baseline，单次推理 <100ms 快两个数量级。
  • Weakness: 训练与评测共用同一合成扰动管线，缺乏真实配对编辑数据交叉验证，核心机制（SeqTag/IterEdit/TagFill）分别源自 NLP 已有方法（GECToR/Levenshtein Transformer/Felix），机制层面新颖性有限。
  • Full review: Claude Code 全文七公理审稿

Music creation is fundamentally a process of revision. Yet symbolic music generation remains dominated by paradigms that produce complete sequences from scratch, with limited support for selective modification. Edit-based methods have proven effective for text transformation tasks, but remain largely unexplored for symbolic music. We trace this absence to the representational level: conventional event-based music encodings lack the structural properties required by explicit music editing. In contrast, the BEAT encoding, a beat-grid-anchored representation originally designed for autoregressive generation, possesses structural properties amenable to editing. We propose BeatEdit, the first framework for symbolic music generation based on explicit edit operations, recasting generation as producing new content by editing a draft rather than synthesizing from scratch. BeatEdit comprises three complementary mechanisms along an axis of increasing edit density: per-token sequence tagging for error correction, iterative refinement for accompaniment editing, and tag-then-fill for segment completion. All these mechanisms share a single encoding and pre-trained backbone, achieving higher precision and perceptual quality than autoregressive and diffusion methods across all three tasks, while remaining efficient, with single-pass inference completing in under 100 ms. Cross-encoding evaluation further reveals that encoding design substantially influences editing effectiveness, with notable encoding-method interaction effects. Code is available at https://github.com/Haoyu-Gu/BeatEdit-code

3. Difference-Driven Gating: Adaptive Feature Fusion for U-Net Decoder

Authors: Kai Li, Xuechao Zou, Jiashen Fu, Zijun Yan, Xintong Wang et al.

Categories: cs.CV, cs.SD, stat.ML | 15 pages, 13 figures Score: 7.05/10 (Obj:9 Id:6 Ind:8 Comp:8 Eff:5 Nov:7 Rep:6)

  • Strength: 提出差异驱动门控范式（FDG/EDG），在同一 backbone 受控比较中一致超越 4 种已有融合方法，EDG 在 Synapse 上 +2.98 pp DSC、语音分离 +0.9-1.4 dB SI-SDRi，计算开销可忽略，消融实验证明 Shannon entropy 优于所有替代聚合函数。
  • Weakness: 声称 SOTA 但 Synapse 82.96% DSC 远低于当前 SOTA（86-87%），遗漏 SegMamba/nnU-Net 等强 baseline，代码未公开，predictive coding 动机与实际机制脱节。
  • Full review: Claude Code 全文七公理审稿

The U-Net style models have been widely used in many applications. A critical step in these models is to reconstruct the lower-level features using a top-down decoder. This reconstruction requires precise fusion of high-level semantics and low-level details. Existing attention-based fusion methods typically derive attention weights from the top-down decoder features (global) alone or the correlation between the top-down decoder features and the bottom-up encoder features (local), then modulate the encoder features using these weights. In this work, we explore a different paradigm: deriving attention weights from the difference between the two feature streams. To this end, we propose two difference-based gating approaches: Feature-difference gating (FDG), which directly uses the absolute difference between global and local features to generate adaptive gating maps, and Entropy-difference gating (EDG), which measures the representational certainty of each stream via information entropy and uses their signed entropy difference to derive the attention weights. Both methods produce coupled gating maps that simultaneously modulate the global and local features. Experiments on different tasks including medical image segmentation, remote sensing image cloud removal and speech separation showed that both methods outperformed existing attention-based fusion methods, and EDG performed better. The results suggested a new paradigm for multi-scale feature fusion in the U-Net style structures.

4. Where Speech Enhancement Hurts Recognition: An Inference Time Polar Projection Diagnosis

Authors: Mingyue Huo, Yuheng Zhang, Hao Zhang

Categories: eess.AS Score: 6.74/10 (Obj:8 Id:8 Ind:9 Comp:7 Eff:6 Nov:6 Rep:8)

  • Strength: 极投影框架以两个标量参数($\alpha$,$\gamma$)将SE-ASR mismatch从定性解释转化为可测量WER gap，在2个增强器×2个识别器上一致发现”magnitude is operative, phase is unhelpful, recognizer-dependent optimum”，统计检验严格(paired bootstrap $B=10^4$)。
  • Weakness: 仅在VoiceBank+DEMAND单一数据集上评估极投影分析，且”phase unhelpful”强结论仅基于2个增强器的估计相位，未与最直接的对比方法(Ochiai et al. 2024 error decomposition)做定量实验对比。
  • Full review: Claude Code 全文七公理审稿

Speech enhancement (SE) can substantially improve perceptual quality, yet enhanced speech does not necessarily improve automatic speech recognition (ASR). Existing remedies, such as retraining the enhancer jointly with recognizer or interpolating enhanced speech with the noisy input, can mitigate this mismatch, but common explanations such as artifacts and over-suppression remain qualitative and do not localize which enhancement component harms recognition. We propose inference time polar projection, a diagnosis for STFT domain enhancement. Given a mask $M=Ae^{jφ}$, polar projection forms $M_{α,γ}=A^αe^{jγφ}$, where $α$ controls magnitude strength and $γ$ controls phase correction. Sweeping these controls on frozen SE and ASR models turns ASR degradation into measurable magnitude and phase effects. Our projection analysis shows that magnitude strength is the operative axis, while estimated phase correction provides no recognition benefit. The optimal magnitude strength is recognizer dependent: waveform-input wav2vec2.0 favors strong correction, whereas log-Mel-input, noise-robust Whisper prefers weaker correction. Finally, the projection provides a simple mitigation for any SE front end in the STFT mask domain, without retraining either the enhancer or the recognizer, making it directly useful for voice assistants and agents that rely on enhanced speech.

5. Tight-Frame Reconstruction for Acoustic Intensity Estimation Using Cardioid Microphone Pairs

Authors: Akira Omoto

Categories: eess.AS | Submitted to Acoustical Science and Technology Score: 6.30/10 (Obj:8 Id:6 Ind:8 Comp:8 Eff:5 Nov:6 Rep:5)

  • Strength: 提出球谐泄漏系数 Gn 和有效泄漏指标 Λ(ω)，将麦克风指向性误差与阵列几何配置的抑制能力结合为物理可解释的预测框架，仿真显示 TF24 在 16kHz 实测指向性下实现数度级方向估计误差（TF6 同条件下达 8°）
  • Weakness: 全部结论基于仿真，无原型实验验证；未与 P-P 方法进行直接数值对比；仅使用单一麦克风型号（DPA2012）；无代码/数据开源
  • Full review: Claude Code 全文七公理审稿

This paper investigates acoustic intensity estimation using pairs of cardioid microphones based on the cardioid-cardioid (C-C) method. Unlike conventional pressure-difference techniques, the C-C method is intrinsically less sensitive to the relationship between microphone spacing and acoustic wavelength. However, practical microphones inevitably deviate from ideal cardioid directivity, producing direction-dependent estimation errors. To improve robustness against such errors, a measurement framework based on spherical tight-frame microphone configurations is proposed. Directional intensity components measured along multiple axes are combined to reconstruct the three-dimensional acoustic intensity vector. Furthermore, directivity errors are represented using Legendre polynomial and spherical harmonic expansions, and a geometry-dependent leakage metric is introduced to quantify the error-suppression capability of different microphone arrangements. Theoretical analysis and numerical simulations demonstrate that tight-frame configurations effectively suppress direction-dependent errors through geometric averaging. The proposed leakage metric successfully predicts the influence of microphone directivity imperfections on the reconstructed intensity vector. The results further indicate that accurate wide-band acoustic-intensity estimation can be achieved even with relatively large microphone spacings, which are generally impractical in conventional pressure-difference approaches. % The proposed framework provides a physically interpretable and practically useful approach for acoustic intensity measurement using directional microphone arrays.

6. Teaching Speech Enhancement Models to Sing: Domain Adaptation from Speech Enhancement to Singing Voice Separation

Authors: Paul A. Bereuter, Mark D. Plumbley, Alois Sontacchi

Categories: cs.SD, eess.AS | Accepted for presentation at the International Workshop on Acoustic Signal Enhancement (IWAENC) 2026 Score: 6.16/10 (Obj:8 Id:5 Ind:8 Comp:5 Eff:6 Nov:5 Rep:8)

  • Strength: SE pretraining 迁移到 SVS 在数据稀缺场景下有效，LoRA 仅用 6-12% 额外参数即可保留 90%+ 的 full fine-tuning SVS 性能同时避免 catastrophic forgetting，且代码和数据完全开源。
  • Weakness: 绝对性能未达 SOTA（比 Mel-RoFormer (L) 低约 2 dB SDR），两个模型家族的 pretraining 数据量差异（700h vs 87h）削弱了跨架构泛化结论，且缺乏对 SE→SVS 迁移为何有效的机制分析。
  • Full review: Claude Code 全文七公理审稿

State-of-the-art speech enhancement models benefit from large-scale labeled datasets, whereas singing voice separation models suffer from limited available training data. To address this limitation, we formulate singing voice separation as domain adaptation from speech enhancement to singing voice separation. We investigate two fine-tuning strategies: full fine-tuning and parameter-efficient fine-tuning using Low-Rank Adaptation (LoRA) on a discriminative and a generative model. Models with either adaptation strategy outperform the same architectures trained from scratch by 0.29-1.8 dB in Signal-to-Distortion-Ratio. Full fine-tuning yields the highest singing voice separation performance, but catastrophic forgetting degrades speech enhancement performance. LoRA fine-tuning achieves competitive singing voice separation performance while preserving the original speech enhancement capability with only 6-12% additional parameters compared to the base speech enhancement model. Furthermore, the generative model shows improved generalization to an unseen test set. The results demonstrate that adapting pretrained speech enhancement models is an effective strategy for training singing voice separation models in data-scarce scenarios.

7. The SonicAGI System for the REAL-TSE Challenge

Authors: Kai Li, Wendi Sang, Jintao Cheng, Xiaolin Hu

Categories: cs.SD Score: 6.16/10 (Obj:9 Id:5 Ind:6 Comp:5 Eff:8 Nov:5 Rep:5)

  • Strength: 在 REAL-TSE Challenge 两个赛道全面超过官方 BSRNN baselines（Track 1 排名第 2/28，Track 2 排名第 5/35），bounded lookahead 模块有效解决了 iterative separator 中 lookahead 累积问题，96ms 总延迟满足在线约束。
  • Weakness: 核心架构创新（lookahead 模块、magnitude-domain fusion）缺乏直接 ablation（有 vs 无），系统由多个已有模块组合而成且每个组件的必要性未被独立证明，代码未提及开源。
  • Full review: Claude Code 全文七公理审稿

Real-world target speaker extraction (TSE) remains challenging because target speech, interference, and enrollment are recorded under mismatched acoustic conditions with reverberation, noise, and irregular conversational overlap. This paper describes the SonicAGI submission to the REAL-TSE Challenge (IEEE SLT 2026). We take a data-centric approach that combines fully simulated mixtures from clean speech with real meeting overlaps, and use a frozen offline enhancer to provide a denoised mirror of real targets for auxiliary supervision. For the online track, we introduce SwiftNet-Lookahead, which inserts a single bounded-lookahead module before a strictly causal iterative separator and keeps the total system latency at 96 ms. For the offline track, we use a frame-level enrollment cross-attention USEF-TFGridNet with a magnitude-domain fusion stage that trades off perceptual quality and speaker fidelity. In the official evaluation, SwiftNet-Lookahead ranks second in Track~1 and USEF-TFGridNet ranks fifth in Track~2, both exceeding the challenge baselines. These results suggest that real-data-oriented training and track-specific modeling are effective for conversational TSE.

8. CHARM: Charge Calibration and Acoustic Rescue for LLM-based Multimodal Sarcasm Detection

Authors: Qiyang Sun, Yi Chang, Yupei Li, Xi Shao, Zixing Zhang et al.

Categories: cs.SD | under review Score: 5.74/10 (Obj:8 Id:5 Ind:7 Comp:5 Eff:6 Nov:4 Rep:7)

  • Strength: BiCAL 在 MUStARD 上达到零样本文本 Macro-F1 0.787，ALFR 对弱 backbone 提升达 +0.382，cross-cultural prosodic decoupling 发现具有可迁移洞察价值。
  • Weakness: 核心方法与已有 prompt bias calibration（He et al. 2024）高度重叠且未引用对比，缺少 ensembling baseline 隔离校准效应，仅 2 个数据集支撑 “cross-lingual” 声称，遗漏 Commander-GPT 等近期强 baseline。
  • Full review: Claude Code 全文七公理审稿

Sarcasm detection, the identification of discrepancies between literal and intended meaning, is a fundamental task in affective computing. However, zero-shot instruction-tuned Large Language Models (LLMs) systematically over-predict the positive (sarcastic) class across the entire capability spectrum, while the prosodic cues humans rely on remain underexploited and transfer unevenly across languages. We introduce CHARM (Charge Calibration and Acoustic Rescue for Multimodal Sarcasm Detection), a training-free framework that couples two modules. Bidirectional Charge Calibration (BiCAL) steers the LLM toward opposing sarcastic and literal verdicts along a symmetric axis of charged prompts; the induced directional biases cancel by construction, and a simple aggregation recovers an unbiased pragmatic signal. Acoustic Late-Fusion Rescue (ALFR) then fuses the calibrated votes with prosodic descriptors and LLM-generated auditory-perception probes through a shallow classifier, actively down-weighting saturated text votes in favour of acoustic evidence. Without fine-tuning any backbone, BiCAL attains the highest reported zero-shot text-only Macro-F1 of 0.787 on MUStARD, while ALFR lifts weak backbones by up to +0.382 Macro-F1 on CMMA. A Stouffer meta-analysis confirms statistical significance on MUStARD and CMMA (Z = 13.89 and Z = 34.64, respectively; p < 10^-43). Our analysis further uncovers a cross-cultural prosodic decoupling: low-level acoustics fail to transfer across languages, whereas high-level perceptual abstractions remain robust. Together, these components yield an explainable, cross-lingual multimodal detector.

9. Simple Features and Honest Calibration for Ambivalence and Hesitancy Recognition in Video

Authors: Vikas Kumar, Aditya Mishra, Haroon R. Lone

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

  • Strength: ASR-erased time 是一个新颖且 clever 的 feature reframing（从 ASR chunk timestamp gaps 中恢复 16 维 hesitation 信号，AP 0.718，与其他成员相关性仅 0.11-0.36），calibration study 诚实展示了小数据上 validation-tuned threshold 的过拟合（0.741→0.690），代码完全开源且 determini。
  • Weakness: 论文声称”above the previous edition’s winner”(0.694)但遗漏了已发表的更强基线 BROTHER (0.7465 macro-F1)，AMF 是标准 gated fusion 的命名膨胀且在 holdout 上 F1 反而下降，ASR-erased time 在最终系统中未能提升 F1（论文自承认”raises AP but not F1”）。
  • Full review: Claude Code 全文七公理审稿

We address ambivalence and hesitancy (A/H) recognition in the ABAW 2026 BAH Challenge: given a short interview video, predict whether the person shows signs of A/H. Our system combines affect-specialised text, audio, and visual representations with a small set of readable linguistic hesitation cues, fused by a reliability gate we call Affective Marker Fusion (AMF), and finished with a simple AP-weighted ensemble at a fixed decision threshold. We also introduce \emph{ASR-erased time}: speech recognisers delete fillers and hesitation pauses from the transcript, but the chunk timestamps keep the time those events took, and sixteen features built from these gaps form the strongest and most independent non-verbal channel we measured (AP $0.718$, correlation $0.11$–$0.36$ with all other members). Across controlled experiments we find three things: cross-modal conflict design does not reliably help on BAH; language is by far the strongest channel while affect-specialised audio is a useful second; and calibration matters more than architecture. Fitting ensemble weights and a threshold on the small validation split overfits: it scores $0.741$ macro-F1 on validation but only $0.690$ on the untouched test set. AP-weighting at a fixed threshold instead reaches $\mathbf{0.731}$ on test.

10. Qwen-Audio-VAE Technical Report

Authors: Ziyue Jiang, Dake Guo, Zekai Zhang, Hangrui Hu, Ting He et al.

Categories: eess.AS Score: 5.50/10 (Obj:8 Id:5 Ind:8 Comp:4 Eff:6 Nov:4 Rep:2)

  • Strength: 在 12.5 Hz 极低帧率下实现跨语音/音乐/声音的可用重建（LibriSpeech PESQ 3.975, STOI 0.980），编码 32 分钟音频仅需 541ms，三步编码器加速实现 3.62× 提速且重建质量几乎无损。
  • Weakness: 核心架构为已有组件（DAC 编码器 + VAE + HiFi-GAN 判别器 + window Transformer）的工程组合，缺少判别器/数据组成消融和同帧率公平对比，训练数据 5M 小时专有且代码/模型均未开源，复现性严重不足。
  • Full review: Claude Code 全文七公理审稿

We introduce \textbf{Qwen-Audio-VAE}, a suite of low-bitrate, fast-encoding continuous audio autoencoders designed for scalable general audio generation. The model is built around a simple but important principle: an audio VAE should not only reconstruct diverse audio with high fidelity, but also produce compact latent representations fast enough to support large-scale text-to-audio training. Qwen-Audio-VAE combines a causal encoder-decoder, window Transformer blocks, and multi-discriminator training to achieve a strong balance between reconstruction quality and compression rate. The model is trained at scale on 5 million hours of multi-domain audio, enabling robust reconstruction across heterogeneous acoustic conditions. To further improve computational efficiency, we adopt an asymmetric encoder-decoder backbone and introduce latency-aware encoder pruning to maximize encoding throughput. Experiments on public speech, music, and sound reconstruction benchmarks show that Qwen-Audio-VAE generalizes well across diverse audio domains and is particularly efficient, requiring only 541 ms to encode 32 minutes of audio. Overall, Qwen-Audio-VAE provides a high-quality, compact, and high-throughput representation backbone for efficient general audio generation.

11. VoxENES 2026: Benchmarking Generalization of Speech Spoofing Detectors Against LLM-Era TTS and Voice Conversion

Authors: Aastha Sharma, Guangjing Wang

Categories: cs.SD, cs.AI | Accepted in InterSpeech 2026 Score: 5.47/10 (Obj:8 Id:5 Ind:6 Comp:5 Eff:5 Nov:4 Rep:8)

  • Strength: 覆盖了 7 个 2025 年发布的 LLM 驱动 TTS 和 3 个 VC 系统，公开了 53,628 条双语音频样本，成功揭示 8 个检测器中 5 个 EER≥47%、最佳仅 28.98% EER 的严重泛化失败。
  • Weakness: 因果解释仅为 post-hoc 假设（”brittle artifacts”），无控制实验隔离变量；8 个 baseline 均为 2024 年及之前模型，缺少新检测方法评估；每种合成方法仅 200-500 条原始样本，统计可靠性不足；VoiceWukong 等前序工作已占据”LLM-era speech spoofing benchmark”框架，增量贡献有限。
  • Full review: Claude Code 全文七公理审稿

Modern LLM-driven text-to-speech (TTS) and voice conversion (VC) systems produce synthetic speech that differs from the generators represented in many legacy spoofing benchmarks. This mismatch creates a temporal generalization gap that can overestimate detector robustness under real-world post-processing conditions. We bridge this gap by introducing VoxENES 2026, a bilingual (English and Spanish) benchmark of 53,628 audio samples generated using 10 contemporary speech synthesis methods and evaluated under 10 standardized post-processing conditions. Using VoxENES 2026, we benchmark eight pretrained detectors without fine-tuning and observe substantial performance degradation: the best model achieves 28.98\% EER overall, while most perform near or below random chance across modern generators and perturbations. Our results highlight the reliance on brittle artifacts in current detectors and establish VoxENES 2026 as a practical testbed for developing robust audio spoofing countermeasures.

12. Anysynth:Zero-Shot Instrument Cloning via In-Context Learning and Asymmetric Hierarchical Guidance

Authors: Chong Jing, Junan Zhang, Jing Yang, Yulun Wu, Fan Fan et al.

Categories: cs.SD | rejected by ISMIR 2026 Score: 5.30/10 (Obj:6 Id:5 Ind:5 Comp:5 Eff:5 Nov:5 Rep:5)

  • Strength: 提出embedding-free的in-context conditioning范式替代CLAP嵌入压缩，通过AnySynth(CLAP)消融隔离证明该机制贡献（PANNs提升5.3%–23.3%），并在四个数据集上全面超越TokenSynth和CTD。
  • Weakness: 仅比较2个外部baseline（遗漏FlowSynth/WaveTransfer/AdaTT/[20]等已发表同类工作），无人类评估，MERT绝对值仅~0.5（GT=1.0）差距巨大，且标题”Any”仅覆盖3类乐器/5秒目标。
  • Full review: Claude Code 全文七公理审稿

Zero-shot instrument cloning aims to render an arbitrary [Target MIDI] sequence with the acoustic identity of an unseen instrument given only a short [Reference Audio, Reference MIDI] pair. Existing methods rely on pre-trained embeddings (e.g., CLAP) that compress the reference audio into a fixed-length vector, discarding fine-grained acoustic cues essential for faithful timbre reconstruction. We present Anysynth, an embedding-free neural synthesizer based on in-context flow matching. By conditioning a Diffusion Transformer (DiT) directly on the uncompressed reference audio and target MIDI, our model allows self-attention to dynamically retrieve acoustic details at generation time. Experiments show that \tool outperforms embedding-based and auto-regressive baselines in audio quality, timbre similarity, and melody adherence. Notably, the model exhibits prompt-length scaling: longer reference prompts yield steadily better timbre fidelity, a property absent in embedding-based systems. To optimize controllability, we further propose Asymmetric Hierarchical CFG, which structurally decouples MIDI and reference-timbre guidance based on their natural semantic-acoustic dependency. This asymmetric formulation avoids gradient conflicts and improves both note accuracy and timbre fidelity, pushing the boundary of expressive, zero-shot instrument cloning. Demo audios are available at https://anysynth-demo.github.io/

13. Qwen-Music Technical Report

Authors: Jin Xu, Kangdi Wang, Ruibin Yuan, Shun Lei, Xiong Wang et al.

Categories: cs.SD Score: 5.20/10 (Obj:8 Id:5 Ind:5 Comp:5 Eff:8 Nov:4 Rep:2)

  • Strength: 在 600 条中英文 prompt 上对 5 个 SOTA 闭源系统取得 13/16 客观指标最佳和全部 A/B 盲测胜率超 50%，外部排行榜（Artificial Analysis）排名第三，系统级集成度高。
  • Weakness: Melody-CoT 与已有工作 MusiCoT（2025.03）高度相似且未引用；无 Melody-CoT 消融实验；无代码/数据/模型开源；训练 reward 与评测指标存在重叠。
  • Full review: Claude Code 全文七公理审稿

In this report, we introduce Qwen-Music, a powerful music generation model capable of producing highly musical and high-fidelity songs with complete vocal singing. Qwen-Music supports two core tasks: Text to Music Generation, which create entirely new songs from text descriptions, lyrics, and musical attributes, and Cover Song Generation, which reinterprets existing songs with different styles and vocal characteristics. Architecturally, Qwen-Music integrates three core components: Qwen-Music-Tokenizer, Qwen-Music-LLM, and Qwen-Music-Render. Qwen-Music-Tokenizer compresses audio into a 25 Hz single-codebook stream of Music Semantic Tokens that preserve semantic and melodic information for LLM prediction. Based on these tokens, Qwen-Music-LLM performs autoregressive music semantic modeling, with a key novelty being a melody-token-based chain-of-thought (Melody-CoT) mechanism that plans melodies before full-song generation, improving creativity, musicality, structural coherence, and reference-audio-based melody cloning. To overcome the fidelity limitations of discrete semantic tokens, Qwen-Music-Render performs generative stereo rendering, enriching acoustic details and producing high-fidelity stereo waveforms. Finally, we train Qwen-Music-LLM on more than 5 million hours of multilingual music data covering hundreds of languages. We first apply quality-aware pre-training curriculum, then use progressive post-training, comprising supervised initialization, offline DPO, and online GSPO, to further improve musicality and instruction-following ability. Across 600 Chinese and English prompts, Qwen-Music achieves state-of-the-art results in 13 of 16 objective musicality and audio-quality metrics. Professional evaluators also prefer Qwen-Music over leading proprietary systems. For cover song generation, Qwen-Music preserves reference melodies more accurately than leading proprietary systems.

14. Evidence Subspace Projection: Measuring How Much Evidence Explains Deepfake Detection in Self-Supervised Speech Models

Authors: Yixuan Xiao, Cheng-Wei Lin, Xin Wang, Yassine El Kheir, Arnab Das et al.

Categories: cs.SD | Accepted to Interspeech 2026 Score: 5.20/10 (Obj:8 Id:5 Ind:5 Comp:8 Eff:5 Nov:5 Rep:8)

  • Strength: 提出基于 neuron activation 子空间投影的量化解释方法，在 2 模型 × 3 训练条件 × 6 数据集 × 9 因子上验证，确认 silence shortcut 并发现训练数据多样性 decorrelates signal-level confounds、PT 后 silence 仍持续等可迁移经验规律。
  • Weakness: 未与已有解释方法（SHAP [4]、time-domain relevancy、CCA）做直接对比，缺少组件必要性 ablation 和 sensitivity analysis（阈值、归一化、quantizer 选择），多数 E∆/erank 绝对值 < 5% 且 70%+ 决策方差来源未解释。
  • Full review: Claude Code 全文七公理审稿

Self-supervised learning (SSL) models are widely used as feature extractors for state-of-the-art audio deepfake detection, but it remains unclear how to directly and quantitatively connect what SSL models capture to detection decisions. To address this gap, we propose Evidence Subspace Projection, a method that represents both evidence factors (e.g., attack category, codec, gender, transmission) and authenticity labels in a shared space constructed from SSL models’ neuron activation patterns. By projecting the decision vector onto each evidence subspace, we obtain a scalar ratio that quantifies the explanatory power of each evidence type. We evaluate SSL models in raw, fine-tuned, and post-trained settings on multiple datasets. The results confirm findings from established studies, validating the proposed method, and reveal new insights into model behavior.

15. Unified Gradient Projection: Language-Balanced Continual Learning for Multilingual Low-Resource ASR

Authors: Ziang Ren, Guodong Lin, Yuchen Ai, Kaize Tan, Wei-Qiang Zhang

Categories: cs.CL, cs.SD, eess.AS | Accepted by Interspeech 2026 Score: 5.16/10 (Obj:8 Id:5 Ind:8 Comp:5 Eff:5 Nov:4 Rep:5)

  • Strength: UGP 在 Whisper-large-v3 上实现 FWER=0.04% 的近零遗忘，AWER=9.80%，并跨 3 个模型规模和极端数据稀缺（5h）场景展示了一致的稳定性-可塑性平衡。
  • Weakness: 核心方法是 A-GEM 投影 + 语言平衡采样 + ER 的工程组合，缺少与 LoRA、gradient surgery 等关键 baseline 的对比，未隔离”语言平衡采样”的独立贡献，且在 Whisper-medium 上 AWER 优于 Standard ER 但作者未诚实讨论。
  • Full review: Claude Code 全文七公理审稿

Large-scale pretrained ASR models such as Whisper exhibit strong multilingual capabilities. However, fine-tuning on low-resource languages often causes catastrophic forgetting. Although continual learning mitigates this issue, existing methods struggle to regulate cross-task interference in multilingual settings, where dominant languages bias optimization. We propose Unified Gradient Projection (UGP), which constrains parameter updates using reference gradients from language-balanced replay in a unified projection space. By equalizing per-language contributions in the projection, UGP reduces dominant-language bias and improves cross-lingual stability. We further show that combining gradient-level projection with data-level replay yields complementary gains in stability and plasticity. Across diverse low-resource language groups and model scales, UGP enables effective adaptation while substantially mitigating forgetting. On Whisper-large-v3, it achieves near-zero average forgetting.

16. StoryTeller: Training-Free Narrative Grounding for Long-Form Audio Description

Authors: Seung Hyun Hahm, Minh T. Dinh, SouYoung Jin

Categories: cs.CV, cs.AI | Accepted to the European Conference on Computer Vision (ECCV) 2026 Score: 5.10/10 (Obj:8 Id:5 Ind:5 Comp:4 Eff:5 Nov:4 Rep:6)

  • Strength: 提出 salience-weighted narrative memory + identity graph + VLM verification 的 training-free AD 框架，在自建的 StoryAD-QA benchmark 的 context-conditioned track 上较 AutoAD-Zero 提升 +11.5pp（90+30 设置），SPICE 指标提升 23.3%。
  • Weakness: 在 AD 领域主指标 CIDEr 上低于 AutoAD-Zero（19.1 vs 22.4, VideoLLaMA2 backbone）和 AutoAD-III（21.4 vs 24.0），与同期工作 Shot-by-Shot/NarrAD 缺少对比，且与 ByteDance 团队 2024 年同名工作 “StoryTeller” 产生命名冲突。
  • Full review: Claude Code 全文七公理审稿

Long-form audio description (AD) requires more than describing visible actions: it must preserve characters, events, relationships, and story context across scenes so that blind and low-vision (BLV) audiences can follow a film. Modern video-language models (VLMs) are effective on short clips, but they often treat each moment independently, producing descriptions that miss who characters are, why events matter, and how the current scene connects to earlier narrative context. We propose StoryTeller, a training-free framework for story-aware long-form AD. Instead of relying only on local visual cues, StoryTeller maintains a verified narrative memory that carries forward story-relevant information across scenes, enabling later descriptions to remain coherent, grounded, and contextually informative. Given only raw video and a movie title, StoryTeller can optionally retrieve public movie metadata to resolve names and story context, while accepting only facts that are supported by the video through semantic filtering and VLM verification. The method requires no subtitles, scripts, AD transcripts, aligned captions, character banks, precomputed face identities, or task-specific fine-tuning. To evaluate whether generated AD preserves narrative information, we introduce StoryAD-QA, a question-answering benchmark that tests whether a language model can answer story-context questions using only the generated descriptions. Experiments on standard AD benchmarks and diverse long-form videos show that StoryTeller consistently improves narrative coherence, factual grounding, and story comprehension over strong baselines in automatic, QA-based, and human evaluations.

17. MusicMark: A Robust Generative Watermarking Framework for Music Generation

Authors: Seohwan Yun, Jeeyoung Yun, Yongjin Kim, Juyeon Lee, Sungwoong Kim

Categories: cs.SD, cs.AI, cs.CR | Submitted to IEEE Transactions on Information Forensics and Security Score: 5.05/10 (Obj:8 Id:5 Ind:6 Comp:5 Eff:5 Nov:4 Rep:6)

  • Strength: 在扩散模型去噪步骤中嵌入水印适配器，在 neural codec re-synthesis 攻击下达到 ~90%+ bit accuracy，显著优于 WavMark/AudioSeal 等 post-hoc 基线（~50-60%），并首次引入 cover-song attack 作为评测场景。
  • Weakness: 未与 GROOT (ACM MM 2024)、GenMark、Latent Watermarking (2409.02915, 直接针对 MusicGen) 等至少 4 项生成式音频水印先行工作比较，”first generative watermarking for music” 的核心 novelty 声明不成立；缺少对联合训练目标各损失项的消融实验。
  • Full review: Claude Code 全文七公理审稿

AI music generation has rapidly advanced alongside commercial platforms, raising the need for reliable watermarking for provenance and attribution. However, existing audio watermarking research has largely focused on speech, and applying speech-oriented methods to music is challenging due to music’s complex structure and rich acoustic texture. Most existing methods are post-hoc, adding imperceptible perturbations after generation rather than embedding watermarks as part of the content. This makes them fragile under transformations and especially vulnerable to neural codec re-synthesis, which can discard imperceptible residual signals. Moreover, since generation and watermarking are decoupled, the watermarking step can be bypassed or omitted, weakening provenance guarantees. To address these issues, we propose MusicMark, which, to the best of our knowledge, is the first generative watermarking framework for music. Specifically, MusicMark embeds watermark messages into the semantic latent space during generation, incorporating the watermark as part of the musical content and ensuring robustness against diverse attacks, particularly neural codec re-synthesis. To this end, we introduce a watermark adapter into a diffusion-based generation model to embed watermark messages across denoising steps. The adapter and detector are trained with a joint objective that preserves fidelity by constraining watermarked latents close to their unwatermarked reference latents, while improving robustness through attack augmentations. Experiments demonstrate that MusicMark substantially outperforms post-hoc baselines across diverse attacks including neural codec re-synthesis, while maintaining comparable generation quality. We further introduce a cover-song attack, converting the singing voice while preserving musical content, and show that MusicMark remains more robust than post-hoc methods.

18. Verifier-Guided Twelve-Tone Composition: A Generate-Verify-Repair Harness for Symbolic Music Generation

Authors: Congren Dai, Danni Zhao, Enyang Liu, Michael Ching Yam, Zhancheng Guo et al.

Categories: cs.AI Score: 4.97/10 (Obj:6 Id:3 Ind:6 Comp:5 Eff:5 Nov:5 Rep:6)

  • Strength: 形式化十二音约束为 Z12 上的谓词系统，并通过 generate-verify-repair harness 将 audited delivery yield 从 13.3% 提升至 48.1%、degeneracy 从 0.13-0.24 降至 ~0.05，在 4 个模型上一致。
  • Weakness: 核心识别失败——harness 使用 638 倍 LLM 调用和 157 倍 tokens，论文自身承认”unmatched compute prevents attributing gains specifically to repair”，无 compute-matched baseline 隔离 verify-repair 机制与暴力重试的贡献。
  • Full review: Claude Code 全文七公理审稿

Large language models can produce superficially legal twelve-tone scores that collapse into degenerate textures. We introduce a neuro-symbolic harness that wraps a language-model proposer in a generate-verify-repair-trace loop with symbolic verification. The complete pipeline improves event-local consistency without claiming whole-piece legality. Across 40 controlled tasks and four paired models, audited delivery yield rises from 13.3% under raw generation to 48.1% with the harness, which explicitly abstains otherwise. The pass rate of a narrower collision and serialisation-consistency check rises from 33.5% to 58.3%, while degeneracy remains near 0.05, including under exploratory adversarial prompting. A blinded evaluation by five experts also shows a descriptive aggregate preference for harness candidates over raw generation in adherence, perceived legality, coherence, and overall quality.

19. Semantic Sampling via Learnable Observation Front Ends

Authors: Yuxuan Liu, Guangming Shi, Pengfei He, Shuai Ma, Xiang Cheng

Categories: eess.AS | 13 pages, 4 figures, 4 tables Score: 4.63/10 (Obj:8 Id:2 Ind:6 Comp:5 Eff:4 Nov:5 Rep:6)

  • Strength: 提出将可学习声学前端用于采样阶段而非重建阶段的问题重构，在 4/2/1 kHz 三种低观测率下所有六个指标均优于未在目标域训练的基线模型（如 4kHz 下 SI-SDR 从 12.60 提升至 15.83 dB）。
  • Weakness: 基线模型（AudioUNet, NU-Wave 2, AudioSR）使用公开预训练参数未在 LibriSpeech 上重训练，而本文方法在 LibriSpeech 上端到端训练，比较不公平导致性能提升归因不可靠，缺少”同数据训练+固定采样”的关键消融基线。
  • Full review: Claude Code 全文七公理审稿

Sampling determines the form of information available to downstream reconstruction systems. Conventional lowrate sampling forms finite-dimensional observations directly from the raw waveform, with the sampling rule mainly guided by bandwidth, sparsity, or fixed signal-level structures. For acoustic signals such as speech, however, reconstruction-relevant information is often expressed through content-related spectral-temporal structures rather than waveform samples alone. This paper proposes semantic sampling via learnable observation front ends, where finite-dimensional observations are generated from learned signal responses instead of directly subsampled waveform points. The proposed front end consists of a semantic feature filterbank, a constrained semantic observation matrix, and a low-rate readout module. The filterbank maps the input waveform into multiple acoustic response channels, the observation matrix combines these responses into a small number of observation channels, and the readout module produces low-rate finite-dimensional samples. A reconstruction network is then used to recover the signal from the resulting observations. Experiments on low-rate speech reconstruction show that, under the same observation budget, the proposed semantic sampling front end provides more informative observations than fixed low-rate sampling and neural restoration methods based on predetermined low-rate waveforms. The improvements in waveform fidelity, spectral consistency, and perceptual quality show that learnable observation front ends preserve more useful information for acoustic signal reconstruction under the same observation budget.

20. Beyond Benchmarks: Exposing the Hidden Crisis in Bangla Hate Speech Detection

Authors: Faria Afrin Tisha, Fariya Tabassum, Hafsa Binte Kibria, Md. Nahiduzzaman, Mominul Ahsan

Categories: cs.CL Score: 4.05/10 (Obj:6 Id:5 Ind:5 Comp:3 Eff:5 Nov:3 Rep:3)

  • Strength: 系统展示 BanglaBERT 在 benchmark（F1=0.91）到真实外部数据（F1=0.753，隐含 HS 仅 0.634）的性能落差，emoji 翻译预处理带来 12% F1 提升，定性识别了 5 类隐含 HS 误分类模式。
  • Weakness: 核心诊断证据仅基于 200 条自建外部集（含 8% 合成样本且不公开），无代码开源、无置信区间、缺少 LLM/BanglaHateBERT baseline 对比，多项发现与已发表工作重叠（Ghosal et al. 2023 已研究 Bengali emoji HS; Fortuna et al. 2021 已研究跨数据集泛化失败）。
  • Full review: Claude Code 全文七公理审稿

The spread of hate speech (HS) across different social media platforms (SMPs) poses a major concern for online safety and ethical moderation. Automatic detection of HS remains a challenging task, especially in under-resourced languages like Bangla, due to cultural context, implicit expressions, and informal linguistic patterns. This study aimed to expose the crisis of Bangla HS detection systems by diagnosing how and why benchmark-trained models fail to identify implicit, context-dependent HS. Six architectures (FastText + CNN, FastText + LSTM, FastText + BiLSTM, BanglaBERT, BanglaBERT + CNN, and BanglaBERT + BiLSTM) were trained on benchmark datasets (about 75,000 posts) and a merged multi-source dataset (about 120,000 posts), then externally validated on an annotated dataset (about 200 posts) collected from Facebook, Twitter, and YouTube, labeled as HS and non-HS, where HS was further categorized as explicit and implicit. BanglaBERT achieved an F1-score of 91.4% on benchmark datasets but declined to 75.3% on the external set and 63.4% for implicit HS involving sarcasm and emojis. The accuracy of FastText + CNN dropped from 78.0% to 51.2% under similar conditions. Emoji-aware preprocessing improved implicit HS detection by up to 12%, whereas emoji removal caused a notable decline in performance (F1: 0.75 to 0.63). Frequent misclassifications in politically charged or satirical comments revealed over-policing risks. This study not only exposes the generalization crisis due to implicit, culturally embedded, and emoji-laden expressions but also underscores the need for developing adaptive, emoji-aware, and culturally grounded frameworks that ensure ethical moderation while preserving freedom of expression. Findings of this study provide insights for researchers, SMPs, and policymakers to design more context-sensitive HS detection systems for low-resource languages.

21. Synchronized Three-Dimensional Vocal-Tract Motion for Speech Synchronization via Joint-Embedding Predictive Architecture Alignment

Authors: Sheng Li, Takahiro Shinozaki

Categories: eess.AS | paper submitted to IEEE-SLT2026 Score: 2.84/10 (Obj:5 Id:2 Ind:2 Comp:2 Eff:2 Nov:4 Rep:4)

  • Strength: 提出了将 TTS carrier 音频与 3D 生物力学声道模型同步的完整 pipeline，在 24 个最小对立对词上实现 8.33% WER 和 0.864 JEPA 对齐分数，并公开了 12 个 GUI 检查录像。
  • Weakness: 全部定量指标仅评测 carrier 音频质量而非 3D 构音运动质量，无任何 baseline 对比或消融实验，构音运动的主观评估尚未完成，24 词诊断集无法支撑系统效用结论。
  • Full review: Claude Code 全文七公理审稿

Modern neural speech systems can generate intelligible waveforms, but they usually hide the physical speech-production state that produced the sound. Conversely, biomechanical vocal-tract models expose articulatory structure, contact behavior, airflow routing, and geometric constraints, but direct physical waveform synthesis remains less robust than modern neural vocoders. A duration-preserving acoustic carrier supplies the listening waveform, while a corrected three-dimensional vocal-tract model supplies synchronized jaw, lip, tongue, velum, laryngeal, oral-airflow, and nasal-airflow motion. A joint-embedding predictive architecture (JEPA)-style representation and a reinforcement learning/cross-entropy method (RL/CEM) trajectory-selection loop align articulatory actions to the acoustic carrier and to physical-plausibility constraints. The evaluation contains 12 3D recordings covering 24 minimal-pair stimuli. On the 24-word set, the carrier obtains good automatic speech recognition (ASR) results (an 8.33\% WER, a 4.17\% CER), a UTMOS score of 3.174, a mean JEPA score of 0.864, and a mean timbre-guard score of 0.947.

22. Casting Everything to Online API Services? A Survey of Integrating Localized Speech Recognition Models in Robotic Systems

Authors: Sheng Li, Jing Li, Felix Schijve, Jun Hu, Emilia Barakova

Categories: cs.RO, eess.AS | accepted in 18th International Conference on Social Robotics (ICSR + ART 2026) Score: 2.63/10 (Obj:5 Id:2 Ind:4 Comp:2 Eff:2 Nov:2 Rep:3)

  • Strength: 论文将 ASR 技术演进与机器人部署策略（onboard/cloud/hybrid）结合在同一框架内讨论，Table 1 提供了三种部署策略的定性比较，覆盖了 7 个真实机器人平台案例。
  • Weakness: 全文无任何量化数据（WER/CER/延迟/资源），核心结论为工程常识（”hybrid 最实用”），文献搜索策略不可复现（作者自认），与已有 ASR 综述和机器人 ASR 综述高度重叠，未提供可操作的部署指导。
  • Full review: Claude Code 全文七公理审稿

Automatic speech recognition (ASR) has become a critical component of modern robotic systems because it is one of the most natural and intuitive ways for humans to interact with robots. A commonly used method is to directly use API services online. But is that all we can do? This article provides an overview of how ASR technologies are integrated into various intelligent robots and machines. We discuss the evolution of speech recognition from established approaches to state-of-the-art deep learning models, such as OpenAI’s Whisper. We also list large-scale datasets and open source toolkits that have been widely used in both industry and academia. We structure the survey around ASR model families, deployment strategies in robotics (especially ROS-based, cloud-based, and hybrid solutions), and several real-world robotic platforms. Finally, we outline the challenges of deploying robust speech recognition in robots and discuss future directions, including multimodal interaction in diverse and dynamic environments. This paper can help social robotics researchers better navigate the emerging domain of language-based natural human-robot interaction.