Daily Papers — 2026-07-31

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

1. Do Music Foundation Models Embed Pitch in Helical Structure?

Authors: Hayato Yagi, Shinnosuke Takamichi, Rin Sato, Keitaro Tanaka, Shigeo Morishima

Categories: cs.SD | Accepted by ISMIR 2026 Score: 7.30/10 (Obj:8 Id:6 Ind:8 Comp:8 Eff:6 Nov:8 Rep:8)

Strength: 首次在 MFM 中间表征中定量证实音高螺旋结构，并通过 137 种人工信号条件下的多元回归将螺旋形成归因于八度等价谐波 (2,4,8 次谐波正贡献，3,5,6,7 次谐波负贡献，统计显著 p<0.05)。
Weakness: 仅分析 2 个生成式 MFM (Jukebox, MusicGen)，未隔离 Transformer 与 tokenizer 对螺旋的贡献，未连接下游任务性能，未涵盖判别式 MFM (MERT, MusicFM)。
Full review: Claude Code 全文七公理审稿

This study analyzes the intermediate representations of music foundation models (MFMs) and reports the geometric structures used to represent pitch information. By inputting isolated musical notes into trained MFMs and analyzing their principal components, we reveal that the representations form a helical structure reflecting the octave periodicity of pitch. Furthermore, we show that the clarity and geometry of this helical structure vary not only across models but also with the acoustic properties of the input signals. Our analysis provides a novel approach for clarifying the internal mechanisms of MFMs.

2. Exploring Efficient Waveform Diffusion Models for Foley Sound Generation

Authors: Runwu Shi, Chang Li, Jiahui Li, Jiang Wang, Yaozhong Kang et al.

Categories: eess.AS Score: 6.50/10 (Obj:8 Id:5 Ind:8 Comp:7 Eff:8 Nov:5 Rep:5)

Strength: DP-U-Net 以 3.26M 参数在 DCASE 上 FAD-V=6.34、E-L1=0.009，全面优于 74M 参数的 T-Foley 和 59M 的 Mamba-Foley，参数效率提升 10-20 倍且主观评测 Quality=3.79/Time Align=4.30 仍领先。
Weakness: 缺少同参数量的纯 CNN U-Net 最简 baseline 和 DP module 整体消融，无法确认双轴 attention 相对于”更小模型”本身的贡献；代码未开源，FSD-Kaggle2018 test split 随机种子未指定。
Full review: Claude Code 全文七公理审稿

Recent advances in diffusion models have enabled high-fidelity Foley sound generation directly in the waveform space. Existing waveform diffusion models primarily rely on time-domain architectures, such as CNN-based U-Nets and DiffWave-style models, or frequency-domain Transformers modeling temporal dependencies. However, these systems are typically built with large model capacities and substantial computational costs, leaving compact and efficient waveform diffusion architectures largely underexplored. In this work, we introduce a Dual-Path (DP) architecture for waveform diffusion that performs dimension-wise self-attention along both subband and frame axes in the time-frequency domain. This DP design enables fine-grained temporal-spectral modeling while maintaining high efficiency. Based on the proposed DP backbone, we develop two variants: DP-DiT and DP-U-Net. Experiments on the DCASE and FSD-Kaggle2018 datasets demonstrate their superior performance. Notably, the 3M parameter variant achieves performance comparable to models with more than 50M parameters. Audio samples are available at https://samplesdemo.github.io/DP-Foley/.

3. Leveraging Beam Search Information for Confidence Estimation in E2E ASR

Authors: Yichen Jia, Hugo Van hamme

Categories: eess.AS | Published on IEEE Open Journal of Signal Processing Presented in ICASSP 2026 Score: 6.26/10 (Obj:8 Id:5 Ind:8 Comp:8 Eff:6 Nov:5 Rep:8)

Strength: SR-CEM 以仅 0.6k/0.4k 参数的两层 MLP，仅用 beam search 输出的 score 和 rank 特征实现架构无关的 ASR 置信度校准，在 in-domain token 级 MCE 4.50% vs softmax 20.04%，跨 4 种架构（hybrid/attention/CTC/RNN-T）和 English/Dutch/noisy/conversational 场景一致有效。
Weakness: 核心 rank+top-K 特征已在 Qiu et al. 2021 中使用，新颖性有限；word-level 判别力（RNN-T ROC 0.768）和 out-of-domain 校准（CV token MCE 19.59%）仍弱；遗漏 Hystoc (ICASSP 2024) 同类 baseline；未测试最简 rank-only heuristic baseline。
Full review: Claude Code 全文七公理审稿

To estimate confidence for end-to-end Automatic Speech Recognition (ASR) systems, recent research has proposed Confidence Estimation Modules that incorporate features from the backbone ASR model. Most existing approaches, however, are architecture-dependent. In this paper, we propose the Score-Rank Confidence Estimation Module (SR-CEM), a lightweight module that leverages beam search information to generate token- and word-level confidence scores. Specifically, SR-CEM constructs features by combining the scores and ranks of tokens within a hypothesis. Experiments show that SR-CEM achieves effective calibration on both in-domain and out-of-domain English data. On the in-domain testset, it attains a Maximum Calibration Error of 4.50% and an Expected Calibration Error of 0.30% at the token level, significantly outperforming softmax confidence (20.04% and 1.75%, respectively). At the word level, SR-CEM achieves 8.17% and 0.35%, compared to 17.91% and 1.67% from softmax confidence. Furthermore, we demonstrate its robustness across hybrid and transducer ASR architectures with different decoding strategies, as well as on Dutch, noisy and conversational speech conditions. Our main finding is that SR-CEM is particularly effective in reducing Maximum Calibration Error, which is critical for reliable downstream use of ASR outputs, while maintaining architecture independence and generality across diverse evaluation conditions.

4. Versatile On-device Adaptation at the Edge by Unifying Few-shot, Zero-shot, Continual, and In-context Learning

Authors: Douwe den Blanken, Martin Lefebvre, Charlotte Frenkel

Categories: cs.LG, cs.AI, eess.AS | 11 pages, 8 figures, 4 tables Score: 6.20/10 (Obj:8 Id:5 Ind:9 Comp:6 Eff:6 Nov:5 Rep:9)

Strength: 首次在单一 SoC（Chameleon）上统一部署 FSL/CL/ZSL/ICL 四种学习场景，跨图像/音频/token 三模态，功耗 9.5 µW–7.83 mW，面积开销仅 0.5%，代码全开源。
Weakness: 各组件均为已有方法（PN/RN/TCN），ZSL 硬件精度仅 60.6%（4-bit 量化致 19.6 点损失），ICL 绝对精度 46.2% 勉强可用，FSL 核心结果来自前作 [25]，ZSL/ICL 缺同类硬件 baseline 无法做公平比较。
Full review: Claude Code 全文七公理审稿

With the ever-increasing pervasiveness of smart edge devices, the demand is growing for applications that can be tailored to users (e.g., custom keyword spotting) or patients (e.g., adaptive health monitoring). Yet, most edge devices rely on fixed inference algorithms and thus cannot learn on-device to personalize predictions. When they can, devices typically support only a specific learning scenario, such as few-shot learning (FSL): going beyond this requires resorting either to another specialized device or to cloud-based retraining, which implies significant energy and latency overheads, a lack of real-time capabilities, and privacy concerns. In this work, we introduce embedder-centric learning (ECL), a framework that unifies four different online learning scenarios: FSL for on-the-fly customization, continual learning (CL) for knowledge accumulation, zero-shot learning (ZSL) for leveraging semantic data, and in-context learning (ICL) for adapting beyond classification. We demonstrate in silicon that ECL can be deployed on resource-constrained devices across four real-world use cases representative of the aforementioned learning scenarios. Our approach establishes a new state-of-the-art performance for FSL character recognition (Omniglot: 96.8% for 5-way 1-shot, 83.3% for 32-way 1-shot), and the first hardware baseline for CL in keyword spotting (NeuroBench keyword FSCIL: 71.8% for 200-way 5-shot). Moreover, we present the first hardware demonstrations of ZSL with semantic data (60.6% for 5-way spoken sentence classification) and ICL (46.2% at the 500th token of RegBench) operating at micro-to-milliwatt power budgets. Therefore, by unifying multiple learning scenarios, we pave the way for smart and versatile devices that can adapt right at the edge, without reliance on the cloud.

5. ParaASR: Multi-Token Prediction for Fast and Long-Context LLM-Based Speech Recognition

Authors: Qingjian Lin, Yuxin Li, Haoyang Zhang, Jun Chen, Yechang Huang et al.

Categories: cs.SD | 14 pages, 3 figures, 4 tables Score: 5.50/10 (Obj:8 Id:5 Ind:6 Comp:5 Eff:8 Nov:4 Rep:3)

Strength: 将多 token 预测与自回归验证相结合，在标准中/英文及长文本基准测试上实现了强劲的 ASR 准确率 (CER 2.97%, WER 3.68%, 长文本 3.70%)，RTF 为 0.0053，每个前向步骤平均接受 5.0/6 个 token。
Weakness: 忽略了直接的前序工作 SpecASR (arXiv 2507.18181, 2025 年 7 月)，该工作已通过 speculative decoding 解决了基于 LLM 的 ASR 加速问题；缺乏代码/模型发布；依赖于 100K 小时未描述的专有训练数据，使得结果不可独立复现。
Full review: Claude Code 全文七公理审稿

Audio-encoder-LLM-decoder architectures have become the dominant paradigm for modern automatic speech recognition (ASR), improving transcription quality through large-scale language modeling. However, the cost of autoregressive decoding scales with decoder size, creating a fundamental trade-off between recognition quality and serving latency. We argue this trade-off is not inherent: unlike open-ended text generation, ASR outputs are strongly anchored to the input speech signal, providing a natural inductive bias toward high-parallelism decoding. Building on this, we introduce ParaASR, an ASR system that leverages Multi-Token Prediction (MTP) to let a 4B LLM decoder emit multiple tokens per forward step. Starting from a publicly available audio-language foundation, the model first establishes a robust autoregressive recognizer and then aligns five future-token branches through a staged optimization recipe. At inference, it proposes a six-token continuation per step and admits only the verified prefix into the transcript, preserving the safety of standard autoregressive decoding. The average accepted length reaches 5.0 out of 6 proposed tokens, confirming that the deterministic structure of speech makes ASR an especially natural setting for multi-token decoding. ParaASR further retains a native 32K-context window and transcribes up to 30 minutes of audio in a single pass. Across diverse benchmarks, it attains average error rates of 2.97%, 3.68%, and 3.70% on Chinese, English, and long-form evaluations, respectively, while reaching a real-time factor (RTF) as low as 0.0053. These results show that decoder scaling, low-latency inference, and long-context transcription need not be competing goals when future-token proposals are anchored by the acoustic signal and guarded by autoregressive verification.

6. M3-DuplexBench: A Multi-Turn, Multilingual, Multidomain Benchmark for Full-Duplex Spoken Dialogue Models

Authors: Ryo Fukuda, Atsushi Ando, Hiroki Kanagawa, Takatomo Kano, Marc Delcroix et al.

Categories: cs.CL | Submitted to SLT 2026 Score: 5.32/10 (Obj:8 Id:5 Ind:6 Comp:5 Eff:5 Nov:5 Rep:4)

Strength: 首个日语 FDSDS benchmark，通过 teacher-forced full-context 扩展至 timing metrics，发现 timing 行为确实 context-dependent（如 Moshi task domain latency 从 2.453s 降至 0.430s），直接反驳 MTR-DuplexBench “timing is context-independent” 假设。
Weakness: 仅评测 4 个 Moshi-based 模型，无 cascaded/非 Moshi 架构，无人类评测锚点，未提及代码/数据公开——作为 benchmark 基础设施的可用性和泛化性严重受限。
Full review: Claude Code 全文七公理审稿

Full-duplex spoken dialogue systems (FDSDSs) can listen while speaking, enabling natural behaviors such as smooth turn-taking, backchannel handling, and user barge-in handling. However, fair comparisons in multi-turn conversations remain a challenge. In addition, existing benchmarks provide limited coverage of languages and dialogue domains. We propose M3-DuplexBench, a multi-turn, multilingual, multidomain benchmark for FDSDSs. M3-DuplexBench supports English and Japanese and covers both casual conversation and multi-turn question answering. In addition, we evaluate models under multiple dialogue context settings, including single-turn, user-only, and teacher-forced full-context settings, to analyze how dialogue history affects model behavior. Experiments with recent FDSDSs reveal model-specific turn-taking characteristics, clear performance gaps across languages and domains, and mixed effects of dialogue context.

7. Stable Autoregressive Speech Generation with Low-Frame-Rate High-Dimensional Continuous Tokens

Authors: Yi Luo, Rongzhi Gu, Jixun Yao

Categories: eess.AS, cs.AI, cs.LG, cs.SD Score: 5.30/10 (Obj:6 Id:5 Ind:8 Comp:5 Eff:5 Nov:5 Rep:5)

Strength: 通过使用多路径信息路由和残差 classifier-free guidance，在连续 AR 语音合成中展示了可测量的长程稳定性（在 5-60 分钟片段内保持 WER 增长率和 speaker drift 指标），优于 MELLE 和 FELLE。
Weakness: 遗漏了近期强连续 AR 基线（CLEAR 在 LibriSpeech 上 WER 为 1.88%，DiTAR），且从未测试过联合设计主张（Locodec + MP-ELD 必须 co-design 的证据），仅有独立组件的消融研究。
Full review: Claude Code 全文七公理审稿

Balancing sequence length, representational capacity, and long-horizon stability is a central problem in autoregressive (AR) speech and audio generation. Representations with higher frame rates or greater capacity can preserve more signal detail, but they also make streaming generation more vulnerable to distribution drift and AR error accumulation. Conversely, shorter and more compressed representations simplify AR modeling, but their limited bandwidth may discard important components and constrain the upper bound of reconstruction fidelity and generation quality. We ask whether a low-frame-rate, high-dimensional, high-bandwidth continuous representation can be co-designed with a streaming generation framework to support robust high-fidelity reconstruction, strong single-token predictability, and superior long-horizon stability. We decompose this goal into two coupled problems: what geometric and statistical properties a high-dimensional representation space should have, and how an AR continuous-token generator should be structured to resist error accumulation. Accordingly, we propose Locodec, a locally encoded codec that shapes its representation space to improve the interpolatability of a lower-dimensional core manifold and the identifiability of the native high-dimensional coordinates, thereby improving the predictability of high-dimensional high-bandwidth tokens. We also propose MP-ELD, a single-token AR flow-matching framework that uses multi-path information routing and residual classifier-free guidance to mitigate error accumulation. Experiments with 8-Hz, 768-dimensional tokens show that our design preserves reconstruction quality, improves single-token predictability, achieves competitive WER, and maintains stable long-form synthesis, without using external SSL/ASR models, pretrained text language models, or post-training stages.

Authors: Ziwei Cheng, Zhenhua Tan, Zhuomin Zhu

Categories: cs.SD, cs.AI | ACM MM2026 ACCEPTED Score: 4.95/10 (Obj:6 Id:5 Ind:8 Comp:3 Eff:5 Nov:3 Rep:5)

Strength: 在 LRS3 噪声条件（SNR -5dB, 11.6% WER）下展示了鲁棒性提升，消融实验覆盖了全部三个组件并包含非对称权重分析。
Weakness: 三个核心组件（multi-turn cross-attention、quality-aware gating、conditional feature enhancement）均为已有技术的重新命名与组合，clean audio 0.68% WER 仅持平已发表 SOTA（AVUR-LLM），且仅在 LRS3 单一数据集上评测，缺少代码开源。
Full review: Claude Code 全文七公理审稿

Audio-visual speech recognition (AVSR) relies on effective fusion of audio and visual modalities, yet existing approaches treat cross-modal interaction as a single-step operation without structured iterative refinement. We present DoubleHelix, a multimodal fusion framework that reformulates fusion as an iterative cross-modal interaction process with adaptive degradation-aware enhancement. The framework comprises three components including ReverseParallelHelix for multi-turn structured interaction with learned alignment constraints, QualitySensor for learning degradation-aware gating signals, and HelixReplication for consistency-guided conditional feature enhancement. Experiments on LRS3 demonstrate that DoubleHelix achieves 0.68% WER on clean audio, outperforming previous best results by 5.6% relative improvement under matched backbone settings. Comprehensive ablation studies validate each component contribution, including targeted analysis of design choices such as asymmetric pathway weighting. The framework shows improved robustness under evaluated babble-noise conditions, achieving 11.6% WER at SNR -5dB.

9. Model-Agnostic Meta-Learning Initialization for Distributed Multichannel Active Noise Control

Authors: Xiaoyi Shen, Junwei Ji, Woon-Seng Gan, Dongyuan Shi, Jun Yang

Categories: eess.AS Score: 3.95/10 (Obj:5 Id:3 Ind:3 Comp:4 Eff:5 Nov:3 Rep:5)

Strength: 在 DMCANC 场景中验证了 MAML 初始化可加速收敛，在 tonal/broadband/real-recorded 三种噪声条件下均观察到比 zero-init IC-DMCANC 更快的 MSE 下降。
Weakness: 无 ablation、无定量指标、训练与测试使用同一声学系统导致泛化 claim 无独立证据支持，且为同一团队第 4 篇 MAML-ANC 论文，方法新颖性增量极低。
Full review: Claude Code 全文七公理审稿

Distributed multichannel active noise control (DMCANC) has emerged as a scalable framework for large-area noise reduction, where multiple nodes operate local single-channel ANC controllers and exchange essential information to achieve global control. A key limitation of existing DMCANC implementations lies in their reliance on zero or random initialization, which leads to slow convergence of adaptive filters and restricts the efficiency of internode collaboration. To address this issue, this paper introduces a model-agnostic meta-learning (MAML) based initialization strategy for DMCANC. By aggregating heterogeneous acoustic characteristics across nodes-ncluding primary and secondary paths-a MAML framework is trained to learn an initialization that generalizes effectively across distributed ANC systems. The MAML initialization is then deployed to all nodes to improve convergence speed under both stationary and time-varying noise conditions. Numerical simulations applied on broadband and real-world noise demonstrate that the proposed algorithms achieves substantially faster convergence and improved noise reduction performance compared with conventional DMCANC, highlighting the potential of MAML initialization as an effective method for large-scale ANC.