Daily Papers — 2026-07-04

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

Authors: SungHun Kim, SeungJun Baek

Categories: cs.CV, cs.AI

| Accepted at ECCV 2026 Score: 6.95/10 (Obj:8 Id:6 Ind:8 Comp:8 Eff:6 Nov:8 Rep:8)

Strength: 中-高 — shallow parallel fusion 替代 deep sequential 是结构性创新，可迁移到其他多模态融合场景；token filtering + Tri-Modal + gating 的组合设计简洁有效。
Weakness: 论文有 ablation（Table 5, 7），分别移除 token filtering、Tri-Modal attention、modality mixture，每个组件均有贡献。
Full review: Claude Code 全文七公理审稿

中文摘要： Q-TriM针对音视觉问答（AVQA）中深层堆叠注意力网络导致跨模态细粒度信息逐层丢失、误差在融合过程中累积的问题，提出了一种以问题为引导的三模态注意力融合方法。该方法在浅层结构中完成文本、视频与音频三模态的联合推理，通过问题语义驱动注意力分配，避免了传统深层串联式注意力带来的信息衰减。实验表明，Q-TriM在保持较高问答准确率的同时显著降低了模型深度与计算开销，验证了浅层、问题引导的多模态融合在AVQA任务上的有效性。

Audio-Visual Question Answering (AVQA) extends classical VQA by requiring joint reasoning over video and synchronized audio. However, many AVQA systems rely on deeply stacked layers of self- and cross attention across text, video, and audio. Such sequential stacking may incur loss of information such as subtle inter-modal cues over the layers, causing errors to accumulate across sequential attention layers during the fusion. We introduce Q-TriM which performs multi-modal fusion in a shallow and

2. TRACE-EVC: Text-Guided Relative Affective Control for Zero-Shot Emotional Voice Conversion

Authors: Zihan Zhang, Shreeram Suresh Chandra, Zongyang Du, Xiutian Zhao, Aurosweta Mahapatra et al.

Categories: eess.AS, eess.SP Score: 6.95/10 (Obj:8 Id:6 Ind:6 Comp:6 Eff:8 Nov:8 Rep:8)

Strength: 提出并验证了 instruction-guided relative EVC 新范式，TRACE-EVC 在 ESD 上 ACCcls=93%、SECS=0.68、nMOS=3.98 全面超越 label/reference-based baseline，同时开源 TRACE-Instruct 数据集（0.54M 指令）和代码。
Weakness: 未与 PromptEVC/ClapFM-EVC/PromptVC 等 prompt-based EVC 做直接比较；EECS 指标与 emotion2vec 训练目标存在循环重叠；rectified flow 相比简单回归的增量贡献未被隔离消融；4 个 auxiliary loss 权重值未报告。
Full review: Claude Code 全文七公理审稿

中文摘要： 这篇论文针对传统情感语音转换（EVC）依赖固定情感标签或参考音频、无法描述情感转换方向与程度的问题，提出了一种基于自然语言指令的相对情感语音转换任务，通过”让语音更平静”或”听起来更自信”等指令定义相对于源语音的情感变化。作者构建了TRACE-EVC框架，利用文本指令引导零样本条件下的相对情感控制，实现对源语音情感状态的灵活调整。该方法在无需目标情感先验的情况下，能够根据自然语言指令精准调节情感强度与方向，为零样本情感语音转换提供了更具表达力和可控性的解决方案。

Traditional emotional voice conversion (EVC) conditions generation on explicit target emotions like labels or references, defining the target affective state but omitting the direction or nature of the transition. We introduce instruction-guided relative emotional voice conversion, a task where natural-language instructions specify source-conditioned affective transformations (e.g., “make the speech slightly calmer” or “sound noticeably more confident”) instead of fixed targets. To support this

3. TokAN: Accent Normalization Using Self-Supervised Speech Tokens

Authors: Qibing Bai, Shuai Wang, Yuhan Du, Bohan Li, Yannan Wang et al.

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

| Submitted to IEEE Transactions on Audio, Speech, and Language Processing (TASLP) Score: 6.16/10 (Obj:8 Id:5 Ind:6 Comp:5 Eff:8 Nov:5 Rep:9)

Strength: 系统完整、模块化清晰、RL post-training 消融干净；但 GRPO-for-AN 是小迁移，架构重设计是工程组合，缺乏机制解释或 synergy 证明。
Weakness: 三条扩展中只有 RL post-training 有干净 ablation（”w/o post-training”：WER 9.23→9.89，ΔPPG 0.2533→0.2589，且诚实报告 SECS 轻微 trade-off）。SFT/pre-training/CTC 也各有消融。但核心扩展 (i)——联合训练 VQ tokenizer——缺乏对会议版 HuBERT/K-Means 的直接对照消融。
Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对口音归一化任务，旨在将非母语（L2）口音语音转换为标准（L1）语音同时保留说话人身份，解决现有方法依赖自然录制的平行L1-L2训练数据或因使用合成目标而质量下降的问题。作者提出TokAN框架，基于自监督离散语音 token 进行口音归一化，这些 token 由L1与L2联合训练的向量量化（VQ）模型提取。该方法通过在离散 token 空间中操作，规避了对平行录音数据的依赖并避免了合成目标带来的质量退化。实验表明TokAN能在保留说话人身份的前提下有效实现口音归一化，提升了转换后语音的质量与自然度。

Accent normalization (AN) seeks to convert non-native (L2) accented speech into standard (L1) speech while preserving speaker identity. The current techniques either require naturally recorded parallel L1-L2 speech for training, or suffer from quality degradation when supervised by synthesized targets. In this paper, we present TokAN, a token-based accent normalization framework that operates on self-supervised discrete speech tokens extracted from a L1-L2 jointly trained vector-quantization (VQ

4. Probing Low-Level Acoustic Attribute Encoding in CLAP Audio Embeddings

Authors: Héctor Martel, Joe Hennessy-Priest, Taemin Cho

Categories: eess.AS, cs.AI, eess.SP

| Accepted to DAFx 2026 Score: 5.79/10 (Obj:8 Id:6 Ind:8 Comp:6 Eff:6 Nov:5 Rep:5)

Strength: 系统性探针研究揭示 CLAP 及 8 个额外音频基础模型中 RT60/LUFS/RP 近似线性编码、SC 需非线性探针的规律，且发现 amplitude-invariant 架构（Wav2Vec2/WavLM-Large/MERT）按构造丢弃 loudness，跨 5 个数据集和 10 个随机种子结果稳定。
Weakness: 缺少与直接信号处理 baseline（如 Schroeder 积分估 RT60、pyloudnorm 算 LUFS）的对比，且无代码发布，同时 Text-based 属性预测仅为定性演示且在训练范围外失效（cathedral 预测 0.89s vs 物理预期 >2s）。
Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对音频基础模型作为通用特征提取器被广泛使用、但其表征内部结构仍缺乏深入理解这一问题，对CLAP音频嵌入的内部结构展开探查研究。作者通过训练复杂度递增的探针模型，分析CLAP嵌入对三个基本感知维度的编码能力，即混响（RT60）、响度（LUFS）以及频谱内容（以频谱质心SC和相对音高RP度量）。研究旨在揭示这些低级声学属性在CLAP表征中的可解码性与编码位置，从而帮助理解音频基础模型所学表征的内部组织结构。该工作为评估音频基础模型对底层声学属性的编码保度提供了系统化的探查框架。

Audio foundation models are widely adopted as general-purpose feature extractors, yet the internal structure of their learned representations remains insufficiently understood. In this work, we analyze CLAP audio embeddings through a probing framework, studying the encoding of three fundamental perceptual dimensions: reverberation (RT60), loudness (LUFS), and spectral content, measured via spectral centroid (SC) and relative pitch (RP). Probes of increasing complexity are trained to predict each

5. EEG-Based Imagined Speech Decoding Using a Hybrid CNN-SNN Architecture

Authors: Fatima Shalhoub, Mariam Al Mawla, Kabalan Chaccour, Iván López-Espejo, Hoda Fares

Categories: cs.SD, cs.HC

| Accepted to IEEE EMBC 2026 Score: 4.89/10 (Obj:7 Id:3 Ind:6 Comp:5 Eff:5 Nov:4 Rep:6)

Strength: 低 — 标准 CNN + 标准 2 层 LIF SNN，无架构创新，无机制解释，无 spike-level 分析；SNN 的”必要性”未被 ablation 证明，方法层面的真实增量很弱。
Weakness: (1) 无 ablation、无 CNN-only / CNN+ANN-head 最简 baseline，把跨论文不可控比较的 10 点增益归因于 SNN，识别证据不成立（公理二 fatal）；(2) 用 validation 子集当 test set，”超越 SOTA” 的声称在官方口径下不成立（公理三、五）；(3) 标准 CNN + 标准 2 层 LIF，跨领域迁移但未证明 SNN 在本任务上不可替代，新颖性属狭义 first（公理。
Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对脑机接口中想象语音解码面临的EEG信号非平稳、低幅值且高度变异的难题，提出一种融合卷积神经网络（CNN）与脉冲神经网络（SNN）的混合架构，旨在借助脉冲驱动的时序动态特性弥补传统机器学习与深度学习方法在事件驱动信息建模上的不足。该架构利用CNN提取空间特征、SNN捕获细粒度时序脉冲模式，从而更充分地刻画想象语音EEG信号中的时空结构。实验结果表明，该混合架构在想象语音解码任务上优于现有方法，验证了将脉冲神经机制引入EEG语音解码的有效性，为严重言语障碍患者的通信恢复提供了新思路。

Imagined speech decoding using EEG signals has emerged as a promising frontier in brain-computer interface (BCI) research, particularly to restore communication for individuals with severe speech impairments. However, decoding imagined speech remains a complex task due to the non-stationary, low-amplitude, and highly variable nature of EEG signals. Existing methods often rely on classical machine learning or deep learning models that fail to exploit spike-based temporal dynamics or event-driven

6. QuaSR: Quality-Aware Sample Reweighting for Pacific Indigenous Speech Recognition

Authors: Yishun Li, Yang Xiao, Gongping Huang, Eun-Jung Holden, Nick Thieberger et al.

Categories: eess.AS

| 6 pages, under peer review Score: 4.30/10 (Obj:5 Id:4 Ind:5 Comp:5 Eff:5 Nov:5 Rep:2)

Strength: 提出将多维度数据质量（声学/转录/对齐）与模型可学习性结合的乘法 gating 重加权方法，在最小语料 Nguna (1.02h) 上实现 WER 12.4% 相对降低（绝对 -4.01）。
Weakness: 四种语言均来自 Vanuatu 且绝对 WER 仍高达 66%（Lelepa），Bislama/Nafsan 改善在标准差内不显著，缺少 curriculum learning baseline 和纯 model-side/data-side-only 消融，无代码开源且变体选择机制不透明。
Full review: Claude Code 全文七公理审稿

中文摘要： 该论文针对太平洋原住民语言这一低资源场景下的自动语音识别（ASR）问题展开研究，指出此类语料普遍存在声学条件异质、转录不一致以及声学-文本对齐可靠性差异大等缺陷，导致标准微调方法对噪声监督信号过于敏感。为此，作者提出了 QuaSR——一种质量感知的样本重加权方法，通过对训练样本按其监督质量动态分配权重，使模型在微调过程中更关注可靠样本而弱化噪声或误导性样本的影响，从而提升 ASR 在低资源语言上的鲁棒性与识别精度。该方法简单有效，便于集成到现有 ASR 训练流程中，为低资源语言的语音识别提供了一种缓解数据质量不均的实用路径。

Training automatic speech recognition (ASR) models for low-resource languages is challenging due to limited data and highly variable supervision quality. In particular, Pacific Indigenous speech corpora often exhibit heterogeneous acoustic conditions, transcript inconsistencies, and varying degrees of acoustic-text alignment reliability, making standard fine-tuning approaches sensitive to noisy or misleading supervision signals. In this work, we propose QuaSR, a simple yet effective weighting fr

7. NouveauVoice: Generating Novel Pseudo Speakers for Voice Anonymization

Authors: Meiying Melissa Chen, Anastasia Kuznetsova, Zhenyu Wang, Zhiyao Duan

Categories: eess.AS, cs.AI

中文摘要： 针对语音合成与转换技术带来的个人隐私风险，现有说话人匿名化系统难以在生成语音间提供足够的身份差异。本文提出NouveauVoice，一种基于分层深度变分自编码器(NVAE)的伪说话人生成框架，作为独立插件模块与FA-Codec和CosyVoice2等最先进的语音转换后端集成，利用可 tractable 采样与ELBO目标合成高表达性、高多样性的伪说话人嵌入。在类似VoicePrivacy挑战的协议下评估，系统在对抗自动说话人验证攻击时达到超过38%的等错误率，实现强身份隐藏。消融实验表明通过控制NVAE层数可在不同匿名化级别间灵活调整，在严格匿名性、伪说话人多样性与语音可懂度、情感表达等下游效用之间取得合理权衡。

Advanced neural technologies in speech synthesis and voice conversion (VC) have introduced severe risks to personal privacy, necessitating robust Speaker Anonymization Systems (SAS). Existing SAS approaches modify voice characteristics in the hand-crafted feature space or speaker embedding space, often struggling to provide sufficient identity variance across generated voices. In this paper, we propose NouveauVoice, a novel pseudo-speaker generation framework based on a Hierarchical Deep Variati