Tag
The paper proposes MOSAIC, a modality-separated architecture with inter-modal cross-attention and modality-partitioned gradient routing, which improves average macro-F1 by 7.4 points across four tasks on the CEJC corpus while reducing backbone parameters by 72% versus fusion-based baselines.
ETC-Bench is the first comprehensive multilingual benchmark for evaluation-side transcript canonicalization, covering 181 phenomenon types across 12 languages, along with a new SENA metric and a taxonomy-guided agent method (TACS) that outperforms baseline systems.
该论文对比了基于WavLM的音频分类器与基于GPT-4o零样本的转录文本分类器在母语识别任务上的表现，发现音频在低水平(A1)更优，而文本在A2至C1阶段配合完整转录时表现更好。
该论文提出一种噪声鲁棒的强制对齐方法，用上采样WLM特征替代MFCC的GMM-HMM对齐器，并显式建模填充词和非词汇语音事件，在Buckeye和FluencyBank上将词边界误差降低19%和17%，大幅减少灾难性错配。
该论文提出了一种接触式多模态筛查框架，利用与 Amazon Alexa 交互过程中采集的视频、音频和文本信号来检测轻度认知障碍（MCI），实验表明多模态融合模型在 Command-Reading 和 Question-Answering 任务上均达到 0.76 的准确率，优于单一模态模型。
AegisVox 提出一种因子化的临床语音隐私框架，为每位说话者分配一致的代理声音并编辑含标识符的语音轮次，在 63.6 小时模拟对话上比较了 TTS 与编解码语音编辑器在替换实现、说话者关联与临床信息保留之间的权衡。
The paper introduces CircuitTree, a tree-compose data structure and greedy density-scaled algorithm for composing many pruned circuits from HuBERT into a syllable transcriber, achieving 64.8% open-syllable accuracy with 99.5% shared circuitry versus a 66.0% non-sharing ceiling.
该论文提出「先澄清、再聚焦」的声明归一化方法，将对话转化为带来源引用和语义标签的简短陈述，使仅由小模型构成的推理管线能以约100倍更低的成本处理数百万条企业客服对话的分析任务。
Shinji Watanabe announces his Google Scholar h-index has reached 100, crediting decades of work on speech processing research and open-source tooling to his students and collaborators. The post highlights landmark works like ESPnet, Deep Clustering, and SUPERB that underpin modern speech recognition.
FD-VAD proposes an ASR-free, streaming semantic endpointer that maps causal audio windows directly to Continue/Stop decisions for full-duplex voice agents, combining a frozen speech encoder with a parameter-efficiently adapted language model and achieving state-of-the-art EOT recall on TurnBench.
This paper proposes a training-free pipeline for pronunciation transcription that integrates lexical resources and acoustic models to achieve low error rates and high efficiency, outperforming baselines including commercial multimodal LLMs.
This paper challenges the assumption that EER is a reliable proxy for human voice similarity in speech generation, demonstrating that perceptual alignment depends more on training objectives than model performance. It proposes embedding geometry with effective dimensionality as a principled metric for evaluating voice similarity.
The tweet announces the start of Interspeech 2026 in Sydney, highlighting the presentation of 2 tutorials and 19 papers on spoken language models and conversational speech recognition.
The paper introduces BanglaTurn, a benchmark corpus and Whisper-based model for end-of-turn detection in Bangla speech, achieving 84.33% accuracy compared to a 69.28% baseline.
This paper introduces an intelligent wake-up system for virtual assistants that uses contextual trigger detection and synthetic conversational data to improve natural interactions. The authors release code, a dataset, and trained models to promote reproducibility.
This paper introduces StreamFraudNet, a weakly supervised incremental model for detecting phone scams from raw speech, achieving a ROC-AUC of 0.9953 and operating in real-time to provide early warnings during calls.
This paper proposes a phonemically comprehensive, ASCII-only romanization scheme for Thai and Lao, designed for cross-lingual correspondence and machine processing to aid language learning and speech processing.
The paper introduces Duplex Cue, an evaluation framework for assessing in-turn adaptation in full-duplex voice agents, comparing human and model responses to overlapping speech.
The paper presents StepAudio3Realtime, an audio-language foundation model for real-time spoken interaction, using a continuous listen-converse-think-act loop with Think-While-Speaking to achieve deep reasoning and low latency, with top-tier performance on benchmarks like MMSU and Full-Duplex Bench.
This paper presents the SEAR system for multilingual speech MCQ, achieving 90.92% accuracy on the MLC-SLM Challenge through segment-evidence-aware routing, supervised fine-tuning, and reinforcement learning.