「小爱同学」到回答只有两秒，中间站着五道流水线工位：唤醒词、VAD、流式 ASR、LLM、TTS，外加最难的一道暗桩——打断处理。本文按延迟账拆解每道工位的职责、模型选型与失败模式，算清为什么端侧语音助手的优化是一张接力棒时刻表。

转写成文字只是「听见」，音频理解的野心是「听懂」：环境声事件、音乐结构、说话人情绪、副语言信息。本文拆解音频语言模型的技术配方（音频编码器 → 语义 token → LLM）、开源代表 Qwen 系与评测基准 MMAU/AudioBench，以及它正在替代的旧管线。

语音对话「说完—听懂—想好—说出」四段链路各带百毫秒级延迟，叠加起来轻易超过人类对话的停顿预算。本文摊开全链路延迟账本，拆解 VAD、流式识别、首 token 与流式合成各环节，并对比级联与端到端两条架构路线。

Connectionist temporal classification (CTC) naturally supports offline and streaming speech recognition with utterance-l…

Agent 工程系统学习第 13 章（前沿专题）：三个高薪高门槛方向。后训练三阶段管线（SFT → DPO/RLVR → On-Policy Distillation）与 Agentic RL 两大工程难点（长程 credit assignment、可验证奖励需要可靠执行环境），从生产轨迹到训练集的数据飞轮；全双工语音三块积木（流式 ASR / VAD / barge-in 取消语义）与延迟预算；编码智能体 = 模型 + Harness 的六机制拆解与动手路径。

阿里Qwen团队发布Qwen-Audio-3.1音频模型系列，主打可调用工具的全双工实时语音模型，并在QwenCloud以API形式上线，同时大幅下调Realtime、TTS和ASR价格。

Palash Debnath 的全本地开源语音工作台 VoiceStudio（当日涨星 +3,274、★43,728、AGPL-3.0）：把 17 个 TTS 与 7 个 ASR 引擎抽象成可插拔引擎层，覆盖克隆/设计/配音/听写/有声书并内置 MCP。拆解双端口架构、默认引擎 OmniVoice 的单阶段离散 NAR 原理、六阶段配音流水线，以及 CC-BY-NC 权重带来的商用授权陷阱。

Sep 23, 2026 Alibaba's AI team Qwen has released Qwen-Audio-3.1, a lineup of five models for speech recognition (ASR), t…

Pruning large pre-trained transformer-based ASR models such as OpenAI's Whisper has seen great adoption, as pruning the …

盘点 2026 年 10 月主流开源 TTS 七个项目（GPT-SoVITS、CosyVoice、F5-TTS、Fish Speech、IndexTTS、Kokoro 等）：机制、音色克隆方式、中文支持与许可证商用限制，附中文效果/实时率/长文本对比表与 F5-TTS 上手示例，兼谈声音克隆的授权与深度伪造合规风险。

RadixAttention 用前缀树跨请求复用 KV Cache，是 SGLang 的招牌。本文讲机制差异，给 OpenAI 兼容服务上手命令与结构化输出、多 LoRA、投机解码现状，并对照 vLLM/Ollama 谈选型。

Backdoor attacks can be implanted in Large Language Models (LLMs) during training, causing unwanted behaviour when a tri…

Sarvam AI has released Saaras V4, the newest generation of its speech recognition model. It covers all 22 scheduled Indi…

📱小米 18 Pro 系列发布，平板、穿戴与三筒洗衣机同场上新 🤖DeepSeek 发新论文，公开 Agent 训练沙箱 DSec 🍎iOS 27.2 Beta 2 加入运动数据限制，可阻止「摇一摇」广告跳转 🚗蔚来 ES9 交付达…

NVIDIA has released Nemotron 3 Diarization, an open-weight speaker diarization model on Hugging Face. It answers one que…

Kyutai has released **Voice of Reason**, 2 open-weight speech-to-speech models that solve math problems out loud. Both s…

模型入海，阶跃走向人群 #欢迎关注爱范儿官方微信公众号：爱范儿（微信号：ifanr），更多精彩内容第一时间为您奉上。 ]

We define OmniVChat (Omni Video Chat) as the task of native audio-visual dialogue between a user and an omni model. In O…

Question answering has advanced rapidly with large language models, but predominantly for high-resource languages, in bo…

Telecom fraud scripts evolve rapidly and are often designed to resemble routine service conversations, creating two key …

9月16日，网易有道「NEXT，AGENT｜有道AI Open Day」在北京举办。]

Google has released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, its most advanced live dialogue models to dat…

· Google 向全体工程师开放 Claude，Gemini 仍是默认模型 · 鸿蒙智行确认问界合作模式调整，赛力斯主导五项业务 · Waymo 计划 2027 年在东京推出全无人出租车 #欢迎关注爱范儿官方微信公众号：爱范儿（微信号：i…

Safety guardrails in open-weight language models can be readily bypassed using Refusal Feature Ablation (RFA), a techniq…

Backdoor poisoning attacks add poisoned examples to otherwise-clean finetuning data, pairing a trigger with a target beh…

Realtime spoken interaction demands deep reasoning, prompt responses, and fluid turn-taking. We present StepAudio 3 Real…

We report a multi-month engineering program to build Sophea, a production bilingual Greek-English automatic speech recog…

Reducing audio-encoder depth lowers the inference cost of speech large language models, but removing complete blocks per…