Human communication weaves speech, language, and nonverbal behaviour into a continuous exchange of meaning and emotion. This keynote traces the way from sensing emotion to building multimodal systems that assess human states and traits, with applications from mental health assessment to emotionally responsive artificial companions, and the vision of multimodal communication intelligence that helps improve lives.

Speaker tracking is the task of separating and following multiple speakers over time, addressing overlapped speech, talker identity, and time-varying speaker count. The proposed segment-online framework performs speaker separation, voice activity detection, and a two-stage sequential organization strategy (overlap-based stitching and memory-based verification), achieving state-of-the-art tracking on LibriCSS and AMI.

Generative AI is transforming creative practice across music, visual art, and writing, yet most current systems leave human creators on the margins, reducing creation to trial-and-error prompting and repeated sampling. Genuine creative work is an iterative process of expressing and refining intention. The path to meaningful machine creativity therefore lies not in fully automated generation, but in symbiotic creativity: human–AI co-creation with the human in the loop, where machine generation and human aesthetic judgment form a continuous feedback loop and the creator remains in control.

自从有了计算机，让计算机听懂人说话就成了人类的梦想。随着ICT的迅猛发展，过去四十年语音技术取得了令人瞩目的成就，今天语音技术已经在人们生活中无所不在。作为一个在语音行业40年的科研工作者，本报告首先将从第一视角回顾80年代中到90年代初国内的部分语音研究历史，然后简单回顾四十年研究的一些成果和感悟，最后会在当前AGI时代，对语音还有哪些可能值得研究的问题做一点探讨。

本报告聚焦从“离线理解”走向“实时交互”的全模态模型演进，介绍 OpenMOSS 团队在视觉理解、通用音频理解、语音生成与实时视频交互方面的研究进展。报告首先梳理多模态大模型的发展脉络与关键挑战，随后围绕 MOSS-VL、MOSS-Audio 和 MOSS-TTS，说明模型架构、训练方法及其在视频理解、音频理解和语音合成中的能力。在此基础上，报告进一步介绍 MOSS Live：该模型面向持续视频流进行实时理解与交互，支持用户在任意时刻提问，根据场景变化即时回应，并在信息不足时保持静默、在新画面到来时及时修正判断，推动视频模型从“看完再回答”转向“边看边交流”。最后展示 MOSS 多模态系列从模型设计、数据训练到推理系统协同优化的技术路径。

As synthetic voices become nearly indistinguishable from human speech, no single technique can reliably prevent their misuse. This talk presents complementary passive and proactive defenses—from deepfake detection toward word localization, and from audio watermarking toward self-recovery watermarking—and shares ideas on what the speech community can do together.

Automatic speech-to-speech translation has long been a dream technology for humanity. Through numerous breakthroughs over decades of research, we have now reached the stage where speech translation services are available on smartphones. However, many challenges remain before we can achieve translation quality comparable to that of a trained simultaneous interpreter. Key issues include how to translate across languages with different word orders, such as English and Japanese, without waiting for the end of a sentence or utterance; how to balance latency and content fidelity; and how to capture the speaker's intent conveyed through intonation.

Agentic intelligence is now a new paradigm and a core component of the frontier in artificial intelligence research. Agentic intelligence transforms the capabilities of humans to agents in an automatic way. Unlike earlier systems that responded to a single instruction prompt and stopped to wait for further orders from users, modern agentic systems generally pursue a final target over many steps across multiple turns in the long context: they plan, call functions and tools, write and execute code, search for information, evaluate their own intermediate results, and revise their approach as a comprehensive workflow.

AAs Artificial Intelligence evolves from single-turn chat interfaces into autonomous, always-on production systems, AI agents are increasingly operating continuously in the background. In this paradigm shift, tokens are no longer just an ephemeral metric for API billing—they have become fundamental production resources that require systematic organization, governance, and dynamic scheduling. When autonomous agents run non-stop and model capabilities and pricing fluctuate rapidly, how can organizations allocate intelligent compute under strict budget and quality constraints?