论文解读
T-Cache: Fast Inference For Masked Generative Transformer-Based TTS Via Prompt-Aware Feature Caching
语音合成 | 9.0/10
实时处理
论文解读
T-Mimi: A Transformer-Based Mimi Decoder for Real-Time On-Phone TTS
语音合成 | 7.0/10
论文解读
Time-Domain Synthesis of Virtual Sound Source Within Personalized Sound Zone using a Linear Loudspeaker Array
空间音频 | 8.0/10
论文解读
Towards Real-Time Generative Speech Restoration with Flow-Matching
语音增强 | 6.0/10
论文解读
UJCodec: An End-to-end Unet-Style Codec for Joint Speech Compression and Enhancement
语音增强 | 7.5/10
论文解读
VChangeCodec: An Ultra Low-Complexity Neural Speech Codec with Built-In Voice Changer for Customized Real-Time Communication
语音转换 语音增强 | 8.0/10
论文解读
WhisperPipe: A Resource-Efficient Streaming Architecture for Real-Time Automatic Speech Recognition
语音识别 | 6.5/10
论文解读
Hallo-Live: Real-Time Streaming Joint Audio-Video Avatar Generation with Asynchronous Dual-Stream and Human-Centric Preference Distillation
音视频 | 8.5/10
论文解读
Opening the Design Space: Two Years of Performance with Intelligent Musical Instruments
音乐生成 | 6.5/10
论文解读
Predictive Directional Selective Fixed-Filter Active Noise Control for Moving Sources via a Convolutional Recurrent Neural Network
声源定位 | 7.5/10
论文解读
RTCFake: Speech Deepfake Detection in Real-Time Communication
语音伪造检测 | 7.0/10
论文解读
Dilated CNNs for Periodic Signal Processing: A Low-Complexity Approach
语音增强 | 6.5/10
论文解读
Full-Duplex Interaction in Spoken Dialogue Systems: A Comprehensive Study from the ICASSP 2026 HumDial Challenge
语音对话系统 | 6.5/10
论文解读
Sema: Semantic Transport for Real-Time Multimodal Agents
实时处理 | 6.5/10
论文解读
Before the Mic: Physical-Layer Voiceprint Anonymization with Acoustic Metamaterials
这篇论文针对在公共场景（如会议、演讲）中，不可信录音设备可能导致声纹泄露且事后无法补救的问题，提出了EchoMask——首个基于声学超材料的物理层实时声纹匿名化系统。其核心方法是在声音到达麦克风前，通过精心设计的被动声学结构对特定低频段（300-700Hz）进行选择性干扰，该频段对说话人识别至关重要
论文解读
Towards Streaming Target Speaker Extraction via Chunk-wise Interleaved Splicing of Autoregressive Language Model
1. **要解决什么问题**：现有基于生成模型（如扩散模型、自回归模型）的目标说话人提取（TSE）方法依赖全局上下文，难以直接用于实时流式场景，强行适配会导致性能严重下降。 2. **方法核心是什么**：提出首个面向流式TSE的自回归（AR）框架，核心是“分块交错拼接范式”。该范式将混合语音分块
论文解读
BEAT: Tokenizing and Generating Symbolic Music by Uniform Temporal Steps
本文针对符号音乐生成中主流的事件序列（event-based）tokenization方法隐含处理时间规律、导致模型需额外学习时间网格的问题，提出了一种名为**BEAT**的新型网格化tokenization框架。其核心思想是将音乐在时间上均匀离散化为“拍”（beat）作为基本单位，将每拍内每个音高
论文解读
MTR-DuplexBench: Towards a Comprehensive Evaluation of Multi-Round Conversations for Full-Duplex Speech Language Models
这篇论文旨在解决当前全双工语音语言模型（FD-SLMs）评测体系的一个关键缺陷：缺乏对多轮、连续对话能力的系统性评估。现有基准多关注单轮交互或特定对话特性（如打断），忽略了模型在多轮语境下维持指令遵循、安全等核心能力的一致性。为此，作者提出了**MTR-DuplexBench**，一个全新的多轮全双
每日研究速递
语音/音乐/音频论文速递 2026-04-22
共分析 21 篇语音/AI 论文
论文解读
Full-Duplex-Bench-v3: Benchmarking Tool Use for Full-Duplex Voice Agents Under Real-World Disfluency
这篇论文针对当前全双工语音代理评估缺乏真实性（依赖合成语音）和任务简单性（单步调用）的问题，提出了**Full-Duplex-Bench-v3 (FDB-v3)** 基准。该基准的核心创新在于使用**100条真实人类录音**（含五种不流畅性注释），在四个任务域中设计了需要**多步API链式调用**的