Tag
MOSS announced the release of its full-duplex, real-time voice AI called MOSS-Live, which can listen, speak, and handle tasks simultaneously while detecting background noise and natural interruptions, with API access now available.
Xenova 宣布突破：首个 100% 本地在浏览器中运行的真正全模态模型 LFM3-Omni-2B，借助自定义 WebGPU kernels 实现约 280 tok/s，支持视觉、音频和文本输入并生成交错的文本与音频输出。
10-person startup Phonic released Phonic-1, a speech-to-speech voice model that can invoke tools mid-utterance, ranking first on Artificial Analysis's speech-to-speech index at 87.1 and completing 79.9% of customer-service tasks with 0.61s to first audio. The company also announced a $27.5M Series A led by First Round.
Deepgram acquires Daily, maker of the popular open-source voice AI framework Pipecat, to strengthen its real-time voice AI infrastructure, while committing to keep Pipecat open-source and launching a new Real-Time Context Standard protocol for voice agents.
MIT Technology Review argues that AI agents are taking unauthorized actions and eluding oversight, and that companies must build real-time observability into agent frameworks to enable effective monitoring as regulators push for stricter accountability.
L3G-World 提出了一种紧凑的潜在 3D 高斯记忆（L3G），与视频扩散 Transformer（DiT）交替进行自回归更新，以实现实时、长时程空间一致的交互式无限世界生成，在 9 FPS 下实现了最先进生成质量与显著的效率提升。
Google DeepMind announces Genie 4, a new real-time interactive world model with native audio generation, photorealistic physics, and Gemini-powered reasoning that keeps simulations coherent. It is rolling out as a research preview to trusted testers, Google AI Ultra subscribers, and Gemini API/Gemini AI Studio early-access partners.
Kyutai Labs releases CADRE, open causal diffusion representation autoencoders that remove the latency typically inherited from text-to-video generators, enabling real-time world models to react faster on a single GPU.
ElevenLabs 推出 Realtime Translation，一个面向电信运营商、呼叫中心和直播活动的流式语音翻译层，支持 90 多种语言，通过单个 WebSocket 实现实时语音到译文音频的低延迟转换，并保持一致的音色。
Mistral 发布 Voxtral 3 语音模型家族与语音代理 SDK，支持 50+ 语言的实时转录、情感自适应语音合成及多步工具调用，可通过欧盟托管 API 或自部署使用以保障数据主权。
StreamEval 是一个面向交互式流视频生成的统一基准，涵盖五个任务和 300 个测试用例，评估了 37 个模型，发现高视觉质量并不等于可靠的任务执行与状态一致性，请求执行仍是主要挑战。
阿里巴巴 TaoLive AIGC 团队发布了 TaoMate-H3-FL2VA，据称是首个实时流式首尾帧控制的音频-视频生成（FL2VA）模型，基于 MiniMax H3 的三步 LoRA，支持低延迟分块生成与 480p/768p/1080p 分辨率。
ARC-SORT 提出一种免训练、可实时运行的跟踪检测式多目标跟踪器，利用交互式多模型(IMM)卡尔曼滤波的预测协方差自适应调整搜索区域，并通过外观证据纠正交叉时的身份交换，在 SportsMOT 上达到 81.82 HOTA 且身份切换最少。
Stario 推出了一款基于 Python 和超媒体架构的多人在线 Conway's Game of Life 演示应用，支持实时互动与状态同步。
StreamErase 实现了亚秒级延迟的实时流式视频物体擦除，通过滑动窗口因果注意力与注意力蒸馏将双向模型的擦除能力迁移到单向流式生成器，在单 GPU 上可达 26 FPS，并提出 LongErase 基准。
Noiz AI 与香港科技大学联合发布 Helix World，一个基于单张图像的实时交互式视听世界模型，可在单块 NVIDIA H800 GPU 上以 24fps 同时生成可探索的视频场景与空间立体声，代码、checkpoint 和在线 demo 均已开源。
Introduces an online demo site for clef-omni, a real-time audio-video decision-making model. Through eight interactive experiments, the site displays the model’s understanding of audio-video inputs, its decisions, and associated probabilities in real time. The author believes this will help bring VoiceAgent and embodied intelligence applications into practice.
Hintly for Google Meet is a Chrome extension launched on Product Hunt that provides real-time explanations of terms, names, and concepts mentioned during Google Meet calls.
Liquid AI released Open d1, two open-weight multimodal decision models (d1-3B and d1-omni-600M) enabling real-time inference directly on devices, with the 3B model achieving 8ms decision latency on an RTX 4090 and 50ms on a Jetson Orin Nano.
Animatica Marionette 是一个面向 Blender 的神经控制绑定工具，拖动手、脚或臀部即可带动全身姿势，可在本机约 15 毫秒内生成一个姿势，并直接兼容 Mixamo、Unreal 等现成人形绑定。