论文解读
把混叠的音频表示拆开：AudioSAE 用稀疏自编码器解剖 Whisper 与 HuBERT
该文在 Whisper 与 HuBERT 每层编码器激活上训练 BatchTopK 稀疏自编码器，用跨种子稳定性、分类探测与消融、转向和脑电相关验证特征，最强证据是转向使虚假语音检出平均下降 70% 而词错误率仅从 5.1% 升至 5.5%，代价是彻底擦除语音概念需移除 19-27% 特征且强转向会破坏识别。
语音活动检测
会议总览
eacl-2026 论文深度解读
共收录 36 篇 eacl-2026 会议论文的 Reader 深度解读
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
关掉麦克风不行，擦掉人声才行：助老监听的隐私防火墙
音频分类 | 5.9/10
论文解读
先别让模型开口：当语音大模型需要学会拒绝
语音识别 | 8.3/10
论文解读
Towards Operational Conversational Intelligence: A Speech Intelligence Framework
说话人日志 | 6.4/10
论文解读
VAD to the Bone: Ultra-Tiny Speech Activity Detection for Edge Deployment
语音活动检测 | 7.4/10
论文解读
Low-Latency Turn-Taking via Context-Aware Preface Generation in a Real-World Dialogue Robot
语音交互 | 6.1/10
论文解读
Transforming Keystroke Noise to Text: Self-Supervised Acoustic Eavesdropping Attacks on Keyboards
语音活动检测 | 6.5/10
论文解读
HeadRoom: Lightweight, Edge-deployable Pipeline for Adaptive Notification Routing
音频事件检测 | 7.2/10
每日研究速递
语音/音乐/音频论文速递 2026-07-11
共分析 1 篇语音/AI 论文
论文解读
Multimodal Voice Activity Projection for Turn-Taking in Social Robots with Voice-Activity-Related Pretrained Encoders
语音活动检测 | 6.7/10
论文解读
Direction of arrival estimation from distant microphone data using single frequency filtering
语音活动检测 | 7.3/10
论文解读
Fast When, Careful Who: Dual-Process Multiparty Turn-Taking with Diffusion Augmentation
语音活动检测 | 5.9/10
论文解读
MuVAP: Multimodal Multiparty Voice Activity Projection for Turn-taking Prediction in the Wild
语音对话系统 | 7.8/10
论文解读
Bypassing Direct Reconstruction: Speech Detection from MEG via Large-Scale Audio Retrieval
语音活动检测 | 7.0/10
论文解读
PersonaKit (PK): A Plug-and-Play Platform for User Testing Diverse Roles in Full-Duplex Dialogue
全双工对话系统评估 | 6.0/10
论文解读
A Toolkit for Detecting Spurious Correlations in Speech Datasets
模型评估 | 7.0/10
论文解读
Automatic Estimation of Speaker Diarization Error Rate Based on Features of Audio Quality and Speaker Discriminability
说话人分离 | 7.5/10
论文解读
Dual Data Scaling for Robust Two-Stage User-Defined Keyword Spotting
语音活动检测 | 7.5/10