论文解读
用 125k 探针在 VAE 潜空间里拨动音高：扩散音乐生成的可控性不在自回归独占
针对 Stable Audio Open 的 VAE 潜空间可解码音高类别这一可验证性质，论文用冻结的微型卷积探针在去噪后 40% 窗口内以 RMS 归一化梯度轻推潜变量，在 27 次配对试验中将旋律一致性从 0.112 提升至 0.274 且未报告音质代价。
测试时自适应
论文解读
无语音却有文本：用门控低秩投影在解码器表示层抑制 Whisper 幻觉
针对 Whisper 在无语音输入上生成流畅幻觉文本的问题，论文用无语音校准集估计解码器幻觉子空间并在推理时做低秩投影以提升无语音概率，门控版本在保留语音识别的同时把非语音幻觉率大幅降低。
论文解读
失配时只听一条语音：用干净先验把增强输出推向高密度区
针对训练与测试噪声失配导致的增强残留噪声，该文用冻结的自回归干净先验对单条测试语音做 KL 散度自适应，在 DNS 等四组数据上主要提升背景抑制，但需早停且初始已干净时收益很小。
论文解读
EM-KalmanNet: Learned Expectation-Maximization for Adaptive Tracking in Partially Known, Block-Wise Time-Varying State-Space Models
声源定位 | 8.8/10
论文解读
Mitigating Over-Suppression in Speech Enhancement via Inference-Time Rethink-and-Refine Correction Module
语音增强 | 6.2/10
论文解读
SCoPE: Training-Free Audio-Visual Event Perception via Sparse Cross-Modal Prior Exchange
音视频理解 | 7.1/10
论文解读
Music Restoration via Latent Operator Optimization and Diffusion Model Priors
音频修复 | 7.0/10
论文解读
Beyond Prompt Adherence: Auditing Attribute-Level Voice Control in Speech Generation
语音合成 | 7.8/10
论文解读
Experience-Calibrated Contrastive Decoding for Mitigating Hallucinations in LM-Based Text-to-Speech
语音合成 | 7.2/10
论文解读
UOT-IR: Structured Routing of High-Polyphony Symbolic Music into Fixed-Budget Representations
音乐理解 | 5.5/10
论文解读
Few-Shot Open-Set Audio Classification via Transductive Prototype Refinement and Class Logit Enhancement
音频分类 | 6.8/10
论文解读
Voice Memory for Agentic Speech Recognition
语音识别 | 8.2/10
论文解读
Probing Speaker Identity Sensitivity in Audio Deepfake Detectors
语音伪造检测 | 6.5/10
论文解读
Where Speech Enhancement Hurts Recognition: An Inference Time Polar Projection Diagnosis
语音识别 | 6.7/10
论文解读
Training-Free Model Selection and Domain-Aware Score Calibration for First-Shot Anomalous Sound Detection
音频事件检测 | 7.3/10
论文解读
video-SALMONN S: Memory-Enhanced Streaming Audio-Visual LLM
音视频问答 | 7.3/10
论文解读
Adaptive Perturbation Selection for Contrastive Audio Decoding
音频理解 | 5.3/10
论文解读
Transductive Zero-Shot Audio Classification with Audio-Language Models
音频分类 | 6.4/10
论文解读
Test-Time Self-Adaptive Conditioning for Stable Audio-Driven Talking-Head Generation
语音合成 | 7.7/10