Daily Papers — 2026-07-06

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

1. RABBiT: Rapidly adaptive BOLD foundation model via brain-tuning for accurate zero-shot and few-shot prediction of speech-elicited responses in the brain

Authors: Omer Moussa, Mariya Toneva

Categories: cs.CL

| Under review at NeurIPS 2026 Score: 7.20/10 (Obj:8 Id:8 Ind:8 Comp:8 Eff:7 Nov:6 Rep:8)

Strength: RABBiT 以仅 12.9M 参数实现零样本超越 SOTA TRIBEv2 和少样本 115K 参数适应（比 ridge 少 3 个数量级），在 324 个未见被试上验证，且消融明确隔离 brain-tuning（去消降 60%）和 TBT 为关键驱动因素。
Weakness: 三个核心组件中 brain-tuning 和 ROI query transformer 分别直接来自作者先前工作和视觉领域迁移，真正的新增量集中在 SID 参数化和整合方式；训练仅 6 个被试且混合数据集反而下降 8%，扩展性未解决。
Full review: Claude Code 全文七公理审稿

中文摘要： 语言理解的大脑响应具有高度的上下文依赖性，且随实验刺激和个体差异而显著变化，因此难以构建能够同时跨刺激与跨被试泛化的计算模型。针对这一问题，作者提出 RABBiT——一个紧凑的音频到 fMRI 编码器作为基础模型，通过”脑调谐”(brain-tuning)在捕捉跨被试共享结构的同时高效适应新被试与新输入。该模型在零样本和少样本设定下能够准确预测语音诱发的大脑 BOLD 响应，为构建可泛化的语言—大脑响应基础模型提供了一种高效范式。

Language understanding in the brain is context-dependent, varying across experimental stimuli and individuals, which makes it difficult to build computational models that generalize across both. This calls for a foundation model of language-evoked brain activity that can capture shared structure while adapting efficiently to new participants and inputs. We introduce RABBiT (Rapidly Adaptive BOLD foundation model via BraIn-Tuning), a compact audio-to-fMRI encoder designed for accurate zero- and f

2. REDDIT: Correcting Model-Generated Timestamp Drift in ASR without Forgetting via Replay-Based Distribution Editing

Authors: Cheng-Kang Chou, Ming-To Chuang, Ke-Han Lu, Chan-Jan Hsu, Hung-yi Lee

Categories: cs.CL, cs.AI, cs.SD Score: 6.95/10 (Obj:9 Id:8 Ind:6 Comp:8 Eff:8 Nov:6 Rep:6)

Strength: 中 — REDDIT 是 LwF/KL-distillation/scheduled-sampling/last-layer-PEFT 的诚实组合，token 级 CE/KL 解耦编辑是真实但中等的机制增量；ablation 严谨，因果识别到位，但组件无原创性突破。
Weakness: 评测存在中度循环风险，但不是 fatal。(1) in-domain gap/long-gap benchmark：reference timestamp 由”拼接 schedule 精确计算”，不依赖模型，独立——这部分干净。(2) OOD 评测（CV-en、ASCEND）MER 用的是公认数据集，独立。
Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对自回归ASR系统中模型生成时间戳在长非语音区间发生漂移的问题展开研究——转写文本可能仍合理，但解码时间轴会逐渐偏离真实音频。作者自建gap与long-gap基准，在15个可输出时间戳的ASR及音频语言模型上系统评估了这一现象，并发现其具有普遍性。在此基础上提出REDDIT方法，通过基于回放的分布编辑在校正漂移的同时避免灾难性遗忘，使模型在修正时间戳偏移时保留原有能力。实验表明该方法能有效缓解时间戳漂移，为时间戳生成式ASR的可靠性提供了新的基准与解决思路。

Modern autoregressive ASR systems can emit timestamps as decoded tokens, enabling timestamped transcription without frame-level aligners or inference-time post-processing. We show that these generated timestamps can drift across long non-speech spans: the transcript may remain plausible, but the decoded time axis drifts away from the audio. We study this non-speech-induced timestamp drift with self-built gap and long-gap benchmarks across 15 evaluated timestamp-producing ASR and audio-language s

3. Unified Audio Intelligence Without Regressing on Text Intelligence

Authors: Zhifeng Kong, Sang-gil Lee, Jaehyeon Kim, Boxin Wang, Zihan Liu et al.

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

| We release the mode at https://huggingface.co/collections/nvidia/Nemotron-Labs-Audex Score: 6.70/10 (Obj:9 Id:6 Ind:6 Comp:6 Eff:8 Nov:5 Rep:9)

Strength: Audex 在 30B MoE 规模上实现了统一音频-文本 LLM，文本 benchmark 仅退化 1-3 个百分点（AIME 2025: 91.2 vs backbone 92.4），而同类 Qwen3-Omni 退化 10-15 个百分点，同时在 ASR (WER 6.82)、TTS (WER 1.70)、TTA (FDopenl3 66.9) 上达到 SOTA，且模型 checkpoint 已开源。
Weakness: 核心架构和训练方法直接继承自同团队先前工作 UALM，增量主要是规模扩展和任务增加，缺乏对音频编码器/codec/投影方式的消融实验，且训练-评测数据潜在重叠未明确排除。
Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对将音频能力融入大语言模型时往往会牺牲原有文本智能这一核心问题，提出了基于Nemotron-Cascade-2-30B-A3B文本MoE模型构建的统一音频-文本大模型Audex（Nemotron-Labs-Audex-30B-A3B）。方法上采用简洁的统一架构，仅使用单个Transformer解码器：音频输入经编码后投影至文本嵌入空间，文本token与量化音频输出token在生成过程中被同等对待，从而实现音频理解、推理与生成的统一处理。该设计在不引入复杂多分支结构的前提下，将音频与语音能力无缝集成到原有文本模型中。其核心目标在于实现统一的音频智能同时不退步文本智能，使模型在获得音频理解与生成能力的同时保持甚至增强原有的文本推理性能。

Audio intelligence involves understanding, reasoning about, and generating both audio and speech. In this work, we introduce Nemotron-Labs-Audex-30B-A3B (Audex), a unified audio-text LLM built on Nemotron-Cascade-2-30B-A3B, a strong text-only MoE LLM. Audex adopts a simple unified design with a single Transformer decoder: audio inputs are encoded and projected into the text embedding space, while text tokens and quantized audio output tokens are treated uniformly during generation. This architec

4. CARD: Cross-component Audio Representation Distillation for Encoder-Free Audio Captioning

Authors: Ganesh Pavan Kartikeya Bharadwaj Kolluri, Yuchen Zhang, Michael Kampouridis, Ravi Shekhar

Categories: cs.SD, cs.CL Score: 6.42/10 (Obj:8 Id:8 Ind:6 Comp:8 Eff:5 Nov:7 Rep:6)

Strength: 中 — cross-component distillation 是真实增量，简洁有效；但仅限 CLAP-HTSAT 一种 teacher，未验证对 CED/EAT 等非 hierarchical teacher 是否仍成立。
Weakness: 效用是最弱环节。绝对值：CARD∗ AudioCaps CIDEr-D 55.4，Clotho 27.5。相对 encoder-based 上界 (SLAM-AAC+CLAP+LoRA 66.4 / 39.0) 仍低 11.0 / 11.5 个百分点，gap 显著。
Full review: Claude Code 全文七公理审稿

中文摘要： 这篇论文针对现有音频字幕系统依赖冻结音频编码器所带来的推理开销和特征瓶颈问题，提出了一种无编码器的音频字幕模型CARD。该模型在推理阶段移除音频编码器，仅用一个13.2M参数的投影器将音频信息输入到带有合并LoRA适配器的冻结LLM中。训练阶段采用跨组件音频表征蒸馏（CARD）方法，将预训练的CLAP-HTSAT教师模型的知识蒸馏到轻量投影器中，教师模型在推理时被丢弃。该方法通过跨组件蒸馏有效解决了编码器推理成本高和声学特征固定的问题，实现了高效的音频字幕生成。

（注：由于英文摘要被截断，以上总结基于已提供的信息，可能未涵盖完整实验结果部分。）

Modern automated audio captioning systems pair a frozen audio encoder with a large language model (LLM) via a trainable projector, incurring the encoder’s inference cost and bottlenecking the model through its fixed acoustic features. We present CARD, an encoder-free audio captioning model that removes the encoder at inference: a 13.2M projector feeds a frozen LLM with merged LoRA adapters, while the teacher used to train it is discarded. CARD distills a pretrained audio teacher (CLAP-HTSAT) int

5. DuplexChat: Constructing Speaker-Separated Full-Duplex Dialogue Speech at Scale for Spoken Dialogue Language Modeling

Authors: Wataru Nakata, Yuki Saito, Hiroshi Saruwatari

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

| 4 pages, 1 figures, submitted to SLT demo track Score: 6.42/10 (Obj:8 Id:5 Ind:5 Comp:5 Eff:8 Nov:5 Rep:9)

Strength: 中 — pipeline 是已有模块的合理工程组合，缺少 ablation；真正的方法贡献外溢到作者前作 DialogueSidon。
Weakness: 论文是数据型，识别公理主要看”pipeline 各阶段是否被验证必要、且贡献被正确归因”。
Full review: Claude Code 全文七公理审稿

中文摘要： 全双工口语对话模型需要将每位说话人表示为独立流的可分离对话语音数据进行训练，但现有大规模公开语音语料多为单声道，难以满足此类模型的训练需求。本文针对这一数据缺口，提出了DuplexChat——一个面向全双工口语对话模型训练的开源语料库，以及DuplexChat-Pipe构建流水线。该流水线从公开播客源出发，先按语言筛选播客订阅源，再经过语音活动检测、说话人分离、重叠语音保留与文本对齐等步骤，将单声道混合音频转换为按说话人分离的双流全双工对话语音，从而保留自然的轮次切换与同时讲话现象。由此产出的DuplexChat语料规模可观且完全开源，为口语对话语言模型提供了此前稀缺的真实、多轮、带重叠的全双工训练数据。

Full-duplex spoken dialogue models are trained on conversational speech in which each speaker is represented as a separate stream, but existing large-scale public speech corpora are mostly monaural, making them unsuited for SDLM training. We present DuplexChat, an open-source corpus for full-duplex spoken dialogue models, and DuplexChat-Pipe, a pipeline for constructing speaker-separated full-duplex dialogue speech from public podcast feeds. DuplexChat-Pipe filters language-specific podcast feed

6. Context-Aware ASR for Mandarin Technical Lectures

Authors: Ho-Lam Chung, Yiming Chen, Hung-yi Lee

Categories: cs.SD Score: 6.42/10 (Obj:8 Id:8 Ind:6 Comp:8 Eff:6 Nov:5 Rep:5)

Strength: 中 — 极简、reference-free、无训练，problem reframing 有压缩价值；但各组件强先例已存在，新颖增量集中在”self-retrieved glossary”单一约束。
Weakness: 增量真实但偏小。新颖点：(1) term-centric 评测指标 + 讲座型 term-rich benchmark；(2) reference-free 的 self-retrieved glossary（从 first-pass 输出提取，非外部 list）。两者结合是真增量。
Full review: Claude Code 全文七公理审稿

中文摘要： 技术讲座中普通话与英文术语交织，而这些术语虽承载讲座核心含义，却只占少量字符，传统字符错误率（CER）因此会掩盖术语识别失败的问题。针对此，作者构建了一个术语丰富的普通话AI/ML讲座基准，并定义了直接衡量技术术语识别效果的术语中心化指标。在此基础上，论文提出了一种两遍式、无需参考文本的解码方法，利用讲座上下文信息辅助术语识别。该工作为评估与提升技术讲座中混合语言术语的识别质量提供了新的基准与思路。

Technical lectures mix Mandarin speech with English technical terms. These terms carry the core meaning of the lecture, yet they occupy few characters. Character error rate (CER) therefore hides their recognition failures. We study whether lecture context helps recognize these terms. We build a term-rich Mandarin AI/ML lecture benchmark, and we define term-centric metrics that measure technical-term recognition directly. We then propose a two-pass, reference-free decoding method. The first pass

7. Evaluating the Effect of Linguistic Relatedness on Cross-Lingual Transfer in Large Multilingual Automatic Speech Recognition

Authors: Andrei Florian, Cynthia Jayne Amol, Hope Kerubo Ombaba, Xiaoyu Cui, Boniface Mwau et al.

Categories: cs.CL, cs.AI Score: 6.27/10 (Obj:7 Id:9 Ind:8 Comp:6 Eff:5 Nov:5 Rep:8)

Strength: 中 — 实验设计的 6-factor 控制变量是好的工程实践，但方法本身零增量（作者明示复用 Khare/Nowakowski），没有提出替代机制。
Weakness: [Major] 没有 small-model 对照复现 Khare 2021 / Nowakowski 2023 的正向结果，因此”我们反驳了前人”的 claim 缺少同一 setup 下的直接对照——可能只是目标语数据下限（1h vs <1h）不同导致的。
Full review: Claude Code 全文七公理审稿

中文摘要： 该研究针对低资源非洲语言的自动语音识别（ASR）扩展问题，探讨在大型多语言ASR模型中，利用语言同源性进行跨语言迁移是否仍能带来显著收益。研究者设计了覆盖六个因素、两个非洲语料库和四个大型ASR模型的系统化受控实验，以验证语言同源性对跨语言迁移效果的预测能力。实验结果显示，在目标语言数据极少的情况下，先在相关辅助语言上预适配并未带来实际有意义的迁移提升。这一发现表明，仅靠语言同源性难以可靠预测大型多语言ASR中的跨语言迁移增益，也不能作为向低资源语言扩展此类模型的有效策略。> Extending automatic speech recognition (ASR) to low-resource African languages is constrained by the prohibitive demands of data collection at scale. A promising direction is to leverage linguistic relatedness to enhance cross-lingual transfer from a related auxiliary language to the low-resource target by sequentially adapting on both. Although this strategy has shown meaningful improvements in small ASR models, its effectiveness in large ASR remains unclear. We extend this framework to large m

8. Ranking the Impact of Contextual Specialization in Neural Speech Enhancement

Authors: Peter Leer, Svend Feldt, Zheng-Hua Tan, Jan Østergaard, Jesper Jensen

Categories: eess.AS, cs.SD

| Accepted to ICASSP 2026 Score: 6.10/10 (Obj:8 Id:6 Ind:8 Comp:6 Eff:6 Nov:5 Rep:5)

Strength: 跨8种架构配置系统验证了语音增强中 speaker 特化收益最大（SI-SDR 增益 +0.5-1.0 dB），且 Spk+Ns 联合特化增益近似可加（预测误差仅 -0.04 dB），10k 参数小模型可匹配 100k generalist。
Weakness: 所有实验基于 oracle 上下文信息（seen-speaker/seen-noise），未讨论实际部署中上下文估计误差的影响，且核心 ranking 结论与 Kolbæk et al. (2017) 已有结果高度一致，新颖性增量有限。
Full review: Claude Code 全文七公理审稿

中文摘要： 本文系统研究了神经语音增强系统在利用上下文信息（如说话人身份、噪声类型、性别、语言和信噪比）进行特化时的效果差异，模型规模从约1万参数覆盖到200万至500万参数。通过对通用模型在不同数据子集上进行微调，作者发现特化所带来的语音可懂度和质量提升因上下文类型而异。实验结果表明，针对说话人身份进行特化始终能带来最大的性能增益。该研究为如何选择最有价值的上下文信息以提升语音增强系统性能提供了重要参考。

We systematically investigate neural speech enhancement systems, ranging from very small ($\sim$10\,k parameters) to medium-large ($\sim$2-5\,M parameters), which specialize to acoustic conditions using contextual information such as speaker identity, noise type, speaker gender, spoken language, and SNR. By fine-tuning generalist models on specific data subsets, we find that specializing to a speaker’s identity consistently yields the largest gains in estimated speech intelligibility and quality

9. SynSFX: Multi-Model Sound Effects Synthesis Dataset for Deepfake Detection and Evaluation

Authors: Linxi Li, Yuncong Yu, Qianwei Guo, Liwei Jin, Yechen Wang et al.

Categories: cs.SD, cs.AI

| 7 pages, 1 figures Score: 5.60/10 (Obj:8 Id:6 Ind:5 Comp:8 Eff:5 Nov:5 Rep:6)

Strength: 低-中 — 无新检测方法，数据集设计为工程性增量（shared prompt），非机制创新。
Weakness: (1) 补充 EnvSDD/BEAT2AASIST/RawNet3 baseline；(2) 在 Table II/III 中单独报告 shared-prompt subset 的 EER，兑现诊断设计承诺；(3) 公开 prompt 扩展模板与人工过滤规则；(4) 将 OOD fake 从闭源商业 API 替换为或补充为开源 unseen 生成器（如 Stable Audio Open 新版本、未训练过的 AudioLDM 变体），使。
Full review: Claude Code 全文七公理审稿

中文摘要： 现有音频深度伪造检测器在合成音效上的泛化能力有限，而已有的环境音频数据集（如EnvSDD）在规模和生成来源多样性上仍不足以支撑对孤立音效伪造的深入研究。针对这一空白，作者提出了SynSFX，一个大规模音效合成数据集，涵盖7种流行的文本到音频生成模型。该数据集共包含43374个音频片段，其中26452个为合成片段、16922个为真实片段，旨在为音效深度伪造的检测与评估提供更具代表性和多样性的基准资源。

需要我把这段摘要写入仓库中的某个文件（例如 _papers/2026-07-08.md 或对应的论文条目）吗？

While audio deepfake detection has advanced significantly, representative detectors show limited generalization to synthetic sound effects. Existing environmental audio datasets such as EnvSDD provide important initial resources, but remain limited in scale and generation provenance for studying isolated sound-effect deepfakes. To support this direction, we present SynSFX, a large-scale corpus of 43374 clips (26452 synthetic, 16922 real) spanning 7 popular text-to-audio models.

10. Listen, Think, Transcribe: Continuous Latent Test-Time Scaling for ASR

Authors: Ho Lam Chung, Yiming Chen, Dau-Cheng Lyu, Hsiao-Tsung Hung, Hung-yi Lee

Categories: cs.SD Score: 5.58/10 (Obj:8 Id:6 Ind:8 Comp:5 Eff:5 Nov:5 Rep:6)

Strength: 三件套 stabilization 工程上有效，但每个组件都是已知手法（L2 + scalar 门控 + 残差锚定）；方法压缩价值主要在 minimal-data activation 这条经验规律，而非架构本身。
Weakness: 方法不是任意性堆叠，但压缩价值有限。(1) 三个 stabilization 机制确实必要（Table 7 的 +47/+11/+3 pp 退化说明 bounded delta/anchor/gate 互补，不是冗余装饰）。
Full review: Claude Code 全文七公理审稿

中文摘要： 端到端ASR模型一次性完成转录，解码器无法回头处理困难输入，针对这一局限，本文提出LatentASR，在冻结的ASR骨干网络上引入连续潜变量测试时扩展机制。该方法仅训练两个轻量模块：一个Latent Adapter通过对少量潜变量前缀位置进行有界、稳定的迭代更新来逐步细化表征，一个Value Head用于预测额外计算是否有助于提升效果并在适当时机提前终止循环。整体方案参数高效，可作为即插即用的增强模块叠加在已有的ASR骨干之上，为模型赋予在推理阶段”再想一想”的能力。该机制以连续潜变量而非显式文本推理实现测试时扩展，兼顾了语音序列建模的高效性与迭代求精的表达力。由于所提供摘要末尾被截断，具体在Qwen3-ASR-0.6B骨干上的量化结果未完整呈现，但其核心贡献在于以极小参数代价为冻结ASR模型注入自适应推理深度。

End-to-end ASR models transcribe in a single pass, leaving no room for the decoder to revisit hard inputs. We propose LatentASR, a parameter-efficient method that adds continuous latent test-time scaling to a frozen ASR backbone. Two small trainable modules drive it: a Latent Adapter that iteratively refines a few latent prefix positions through bounded, stabilized updates, and a Value Head that predicts whether extra computation will help and halts the loop early. The Qwen3-ASR-0.6B backbone st

11. SPEARBench: A Benchmark for Naturalness Evaluation in Streaming Speech-to-Speech Language Models

Authors: Thomas Thebaud, Yuzhe Wang, Hao Zhang, Sathvik Manikantan Napa Ugandhar, Ashish Hallur et al.

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

| Corresponding Website: https://thomasthebaud.github.io/SPEAR-benchmark-website/#welcome Score: 5.53/10 (Obj:8 Id:5 Ind:5 Comp:8 Eff:5 Nov:5 Rep:5)

Strength: 中 — benchmark 协议清晰、子指标计算公式可追溯、reference condition 设计合理（人类原始答案），但缺最简 heuristic baseline、未做 stance judge 污染控制、未公开每个模型的具体推理配置。
Weakness: 这是 benchmark 论文而非 method 论文，识别公理要求降级——但仍有缺口。
Full review: Claude Code 全文七公理审稿

中文摘要： 摘要被截断了，我来搜索一下这篇论文以获取完整内容。

Streaming speech-to-speech language models aim to answer spoken queries directly with synthetic speech. However, standard speech and text benchmarks do not capture whether these systems behave naturally in conversations, where timing, turn-taking, prosody, interpersonal stance, language and dialect consistency, and relationship-aware appropriateness jointly shape perceived quality. We introduce SPEARBench, a benchmark for evaluating naturalness in speech-to-speech language models from question-a

12. Towards Robust Uncertainty-Aware Speaker Modeling

Authors: Junjie Li, Yang Xiao, Kong Aik Lee

Categories: cs.SD

| Submitted to SLT2026 Score: 5.16/10 (Obj:8 Id:5 Ind:8 Comp:5 Eff:5 Nov:4 Rep:5)

Strength: 中低 — 两个组件均为已知技巧的组合（hardness-aware scaling + 标准 NLL adaptation），intra-speaker term 单独无效，需 exp 调制才生效，组件必要性证据不足。
Weakness: 两个组件的真实增量都很薄：(1) “把 intra-speaker compactness 嵌入 hardness-aware scale” 这一想法，face recognition 中的 MagFace [18]、ScaleFace [31]、AdaSpeaker [33] 已经分别做过 quality-aware 与 gradient-aware scaling；speaker verification 内部 U3-xi 已用 i。
Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对说话人识别中现有不确定性感知方法在域偏移下不确定性估计不准确且校准失效的问题，提出一个从估计与适应两个视角出发的鲁棒不确定性建模框架。在估计层面，作者提出跨/内说话人感知不确定性Softmax（Inter- and Intra-Speaker-Aware Uncertainty Softmax），将说话人间可分性与说话人内变异性同时纳入不确定性学习，使估计值更可靠地刻画说话人嵌入的置信度；在适应层面，提出不确定性校准域适应（UCDA）框架，以缓解域失配导致的不确定性失准。在域内与跨域基准上的大量实验表明，该方法在提升不确定性可靠性与说话人识别鲁棒性方面均取得了稳定一致的改进。

Speaker embeddings aggregate frame-level acoustic features into compact representations for speaker recognition. Recent uncertainty-aware speaker modeling approaches further characterize the reliability of speaker embeddings by estimating their associated uncertainty. However, existing methods often suffer from inaccurate uncertainty estimation and uncertainty miscalibration under domain shifts. To address these challenges, we propose a robust uncertainty modeling framework from both estimation

13. Semantic Homogenization in Italian Popular Music: A Diachronic Analysis

Authors: Lorenzo Canale, Stefano Scotta, Alberto Messina

Categories: cs.CL Score: 5.00/10 (Obj:8 Id:5 Ind:5 Comp:5 Eff:5 Nov:5 Rep:2)

Strength: 在 75 年圣雷莫音乐节语料上用四种互补方法（全文/分段/主题/词汇相似度）验证了歌词语义同质化趋势，三种嵌入模型与统计方法的交叉一致性（Pearson > 0.85）为结论提供了方法论鲁棒性。
Weakness: 核心发现为对 Parada-Cabaleiro et al. (2024) 英语结论的跨语言确认，无因果识别、无控制语料、无独立人类校验；代码/数据/prompt 均未公开且依赖闭源 LLM（Gemini、OpenAI），基本不可复现。
Full review: Claude Code 全文七公理审稿

中文摘要： 该研究探究意大利流行音乐歌词是否存在语义同质化趋势，以圣雷莫音乐节75届共75首决赛歌曲为语料，考察非英语文化语境下歌词语义多样性是否同样下降。核心贡献是开发了一套灵活高效的研究方法，用于跨时代、跨语言的歌词语义分析与比较。方法上结合自然语言处理与历时（diachronic）分析，对歌词的语义内容、词汇丰富度等维度进行量化追踪。结果显示意大利流行音乐歌词也呈现出语义趋同和多样性下降的迹象，与英语流媒体歌曲的全球性趋势相呼应。这一发现为音乐文化同质化现象提供了跨语言证据，并验证了所提方法在不同文化语境中的适用性。

In recent years, studies have revealed a decline in semantic variety across popular music lyrics, particularly in English-language songs on streaming platforms like Spotify. This research examines whether a similar trend can be observed in a different linguistic and cultural context: the lyrics of all finalist songs from the 75 editions of the Sanremo Music Festival, Italy’s most renowned music competition. What sets this work apart is the development of a flexible and efficient methodology for

14. Streaming Neural Speech Codecs through Time-Invariant Representations

Authors: Kélian Estève, Salima Mhdaffar, Mickael Rouvier, Richard Dufour, Yannick Estève

Categories: cs.CL

| Accepted to SPECOM 2026 Score: 4.95/10 (Obj:5 Id:4 Ind:8 Comp:5 Eff:4 Nov:4 Rep:5)

Strength: 通过五个 probing 任务系统验证了 TiCodec TIRE 表示主要捕获 acoustic scene 和 emotion 信息而非 speaker identity，并实证流式 660ms 块解码可行（MOS 仅降 0.003）。
Weakness: Dual-TIRE 改进幅度极小（ViSQOL +0.028，Sim +0.022，PESQ 反降），未与 EnCodec/DAC/FreeCodec 等标准 codec 对比，且代码未开源。
Full review: Claude Code 全文七公理审稿

中文摘要： 本研究针对神经语音编解码器中时变语音内容与时不变信息耦合建模的问题，探讨TiCodec通过TIRE模块提取的时不变表示所捕获信息的本质及其在流式场景下的适用性。方法上，该工作对TIRE表示编码的语义内容进行系统性分析，并评估其作为解耦因子化表示在流式语音编解码中的有效性。结果显示，时不变表示能够有效分离语音中的静态属性（如说话人身份、信道特征等），从而显著降低帧级模型需处理的信息量。该研究为理解TiCodec解耦机制提供了实证依据，并为流式语音生成系统的设计指明了优化方向。

Neural speech codecs are increasingly used as intermediate representations in codec-based speech generation systems. TiCodec introduces a factorized representation that separates time-varying speech content from time-invariant information through a Time-Invariant Representation Extraction (TIRE) module, potentially reducing the amount of information that must be modeled at the frame-level. In this work, we investigate the nature of the information captured by TIRE representations and their suita

15. ProPS: Prompted Profile Synthesis for Natural Language-Conditioned Speaker Embedding Distributions

Authors: Thomas Thebaud, Junhyeok Lee, Laureano Moro-Velazquez, Jesus Villalba Lopez, Najim Dehak

Categories: eess.AS, cs.AI Score: 4.42/10 (Obj:5 Id:3 Ind:4 Comp:5 Eff:3 Nov:4 Rep:7)

Strength: 中 — 方法本身简洁、可复现、工程上合理；但本质是 SBERT + MDN + profile-GMM-bank 的组合，机制上是软检索，无新机制。
Weakness: 严重缺陷。论文的核心 claim 是”ProPS（prompt-conditioned MDN over x-vector space）能从 NL prompt 生成对应的 x-vector 分布”。但实验设置无法识别”prompt 真的起作用”vs”prompt 仅起路由作用”
Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对现有说话人嵌入提取器仅具描述性、无法从语言描述生成说话人身份的问题，提出了ProPS（Prompted Profile Synthesis）框架，通过自然语言提示（如”三十岁印度口音男性说话人”）条件化生成说话人嵌入分布。该框架将传统的”语音到嵌入”映射转变为”文本到嵌入分布”的生成范式，使下游任务能够直接从语言描述获取符合目标属性的说话人表征。ProPS突破了现有x-vector方法只能从观测语音段提取单一嵌入的局限，为说话人相关的应用提供了更灵活的生成式解决方案。

Speaker embeddings, or x-vectors, are widely used to represent speaker identity and speaker-related attributes, but existing embedding extractors are typically descriptive rather than generative: they map an observed speech segment to an x-vector, which is then used for downstream applications. We introduce ProPS, Prompted Profile Synthesis, a framework for generating distributions of speaker embeddings conditioned on natural language prompts such as “a thirties male speaker with an Indian accen

Authors: Qu Yang, Cakra Wardhana, Tim Ng

Categories: cs.CL Score: 4.00/10 (Obj:8 Id:4 Ind:5 Comp:3 Eff:4 Nov:3 Rep:2)

Strength: 在 SEAME 上利用 178k 小时伪标签数据将 MER 从 19.23%/27.18% 降至 12.88%/18.89%，证明了大规模伪标签数据对 CS-ASR 的有效性。
Weakness: 方法是 Xu et al. 2020 迭代伪标签的直接迁移，缺少迭代 vs 非迭代的关键消融；Baseline 过弱（96.6h vs 178kh 数据量差 1800 倍）；绝对性能落后于 2026 年 SOTA（8.9%/14.2%）；大量私有数据和私有工具使结果不可复现。
Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对汉英码切换（Code-Switching）语音识别中训练数据稀缺的难题，首次将迭代式伪标签（iterative pseudo-labeling）方法引入码切换ASR领域，利用无标注数据提升识别性能。该方法包含三个阶段：伪标签生成、两阶段双语模型训练以及迭代优化，通过逐步精化伪标签质量来持续改进模型。实验表明，该渐进式精化策略能有效挖掘无标注语音数据，显著提升汉英码切换ASR的识别准确率，为低资源码切换场景提供了一种可行的训练范式。

Code-switching (CS), alternating languages within the same utterance, poses significant challenges for automatic speech recognition (ASR) due to limited CS training data. This paper applies an iterative pseudo-labeling training approach to CS-ASR for the first time, demonstrating its effectiveness in leveraging unlabeled data to improve CS-ASR performance. The approach comprises three phases: pseudo-label generation, two-stage bilingual model training, and iterative improvements. It begins by ge

17. Towards Language-Agnostic Speech Inversion

Authors: Saba Tabatabaee, Mark Tiede, Suzanne Boyce, Liran Oren, Carol Espy-Wilson

Categories: eess.AS

| Accepted to be presented at Interspeech 2026 Score: 3.26/10 (Obj:5 Id:2 Ind:5 Comp:4 Eff:4 Nov:2 Rep:4)

Strength: 在英语训练的 SI 系统上验证了跨语言泛化（法语 PPMC 0.83、俄语 0.74），并首次将 VP TV 估计纳入跨语言评估（法语 0.89、俄语 0.82）。
Weakness: 方法是作者前作 [8] 的增量扩展（输出层 +1 unit + 更换 backbone），无 ablation、无外部 baseline 比较、样本极小（法语 4 人/俄语 3 人），且 “first multi-lingual evaluation” 的 novelty claim 被 Hao et al. (Interspeech 2024) 挑战。
Full review: Claude Code 全文七公理审稿

中文摘要： 该论文研究语音反演(SI)系统如何从语音声学信号中恢复特征性时间模式，包括口腔声道变量(舌头和嘴唇收缩)以及声源信息(周期/非周期能量和基频)。作者开发了一个SI系统，可同时估计口腔声道变量和声源信息，旨在构建跨语言无关的语音反演框架。该方法利用语音信号中反映声道构型和声学激励的时序特征进行反演建模。核心目标是通过联合估计多种语音特征提升SI系统的泛化能力，使其不受特定语言限制。主要贡献在于验证了从语音中同时恢复声道构型与声源信息的可行性，并探索了跨语言反演的潜力。

Characteristic timing patterns are reflected in the acoustic speech signal, encompassing both vocal tract configuration and acoustic excitation. Previous studies have demonstrated that speech inversion (SI) systems can recover these timing patterns from speech, including oral tract variables (tongue and lip constrictions) and source information such as periodic and aperiodic energies and fundamental frequency. In this study, we develop an SI system that simultaneously estimates oral tract variab

18. Quantum-Inspired Harmonic Decision Models: A Computational Framework for Music Generation

Authors: Josef Pavlíček, Petra Pavlíčková, Martin Molhanec

Categories: cs.AI, cs.SD

| 17 pages, 3 figures. Preprint. Code and evaluation data available at GitHub Score: 2.74/10 (Obj:5 Id:2 Ind:5 Comp:2 Eff:2 Nov:2 Rep:5)

Strength: 提出了一种将和声化形式化为约束组合优化问题的混合框架，classical optimization 阶段在 2 首曲目上将 chord density 从 4.00 降至 2.53-2.71，chord duration 从 1.00 提升至 1.87-1.98。
Weakness: 无任何外部 baseline 比较、无 ablation 隔离量子启发式组件贡献、仅 2 首极短曲目（8-9 beats）且 50% 案例专家评估为”less natural”，核心”量子”机制实为纯经典加权迭代且未引用至少 4 篇直接相关的量子音乐先行工作。
Full review: Claude Code 全文七公理审稿

中文摘要： 本文针对音乐生成中的和声决策问题，提出了一种受量子计算启发的计算框架，将和声化建模为结构化组合空间中的优化问题，在多重相互作用的音乐约束下对多个候选和弦序列进行评估。该模型将基于干涉的和声化阶段与源于调性和声的经典优化过程相结合，量子启发式组件能够并行考虑多个和声路径，从而在保持调性逻辑的同时拓展搜索的多样性。该框架为符号音乐生成中的和声选择提供了一种统一的计算视角，兼顾了音乐理论的规则性与生成方法的灵活性。

The user provided a paper title and English abstract that was cut off mid-sentence. Let me write a 3-5 sentence Chinese summary based on what’s available.</think>本文针对自动音乐生成中的和声决策问题，提出了一种受量子计算启发的计算框架，将和声化建模为结构化组合空间中的优化问题，在多重相互作用的音乐约束下对候选和弦序列进行评估。方法上，模型将基于干涉机制的和声化阶段与植根于传统调性和声的经典优化过程相结合，量子启发式组件能够并行考虑多条和声路径，从而在满足音乐理论约束的同时高效搜索解空间。该工作为符号音乐生成中的和声选择提供了统一的计算视角，兼顾了音乐理论的规范性与生成方法的灵活性。

This paper introduces a quantum-inspired computational framework for harmonic decision-making in music. The proposed approach formulates harmonization as an optimization problem within a structured combinatorial space, where multiple candidate chord sequences are evaluated under interacting musical constraints. The model combines an interference-based harmonization stage with a classical optimization procedure grounded in tonal harmony. The quantum-inspired component enables the parallel conside

音频 · 语音 · 音乐 · 声学