---
{"schema":1,"date":"2026-09-23","timezone":"Asia/Shanghai","title":"电话语音 Agent 每日观察","revision":1,"publication":"publish","review_status":"generated","summary":["千问AI耳夹耳机从展示走到预约，但关键耳机状态仍未公开。","新论文把“何时主动插播”变成可训练的中断/静默决策。","Qwen Realtime与Bolna补强工具调用和通话断线恢复。"],"events":[{"project":"qwen-ai-clip-earbuds","title":"千问AI耳夹式耳机在云栖大会开启预约，10月13日发售；现场资料称支持面对面翻译、AI听记和语音办事","kind":"industry","event_date":"2026-09-22","publication_date":"2026-09-22","discovered_date":"2026-09-23","evidence":"source-claim","source":"https://www.ithome.com/1/005/647.htm"},{"project":"proactive-audiollm-ism","title":"七日补漏：论文用interrupt/silent两类决策训练AudioLLM判断何时主动提醒并抑制重复打扰，面向可穿戴实时音频","kind":"paper","event_date":"2026-09-18","publication_date":"2026-09-18","discovered_date":"2026-09-23","evidence":"source-claim","source":"https://arxiv.org/abs/2609.21183"},{"project":"qwen-audio-3-1-realtime-plus","title":"七日补漏：阿里云上线Qwen-Audio-3.1-Realtime-Plus，支持全双工语音、函数调用、Web搜索和语音克隆","kind":"release","event_date":"2026-09-20","publication_date":"2026-09-20","discovered_date":"2026-09-23","evidence":"verified","source":"https://www.alibabacloud.com/help/en/model-studio/qwen-audio-3-1-realtime-plus"},{"project":"bolna-0-10-251","title":"Bolna 0.10.251修复Sarvam TTS在通话中WebSocket断开后的接收恢复，减少语音链路半途卡死","kind":"release","event_date":"2026-09-22","publication_date":"2026-09-22","discovered_date":"2026-09-23","evidence":"verified","source":"https://github.com/bolna-ai/bolna/releases/tag/0.10.251"}]}
---
# 电话语音 Agent 日报｜2026-09-23

今天最接近目标的新进展，是千问把 AI 耳夹式耳机从“展会原型”推进到可预约产品；与此同时，一篇七日补漏论文正面研究“助手何时该主动打断”，但完整的“耳机入口→后台长任务→未确认后电话兜底”仍没有现成成品。

1. **千问 AI 耳夹式耳机开始从概念走向可买。** 9 月 22 日云栖大会上，它开启预约并定于 10 月 13 日发售，现场资料称支持面对面翻译、AI 听记和语音办事，方向与“长时间戴着、随口交给 Agent”很接近。可惜公开材料还没说明唤醒词、入耳/摘戴检测、跨第三方播放器压音恢复、文字未读/确认或电话升级；本轮也只核到国内预约信息。[云栖现场](https://www.ithome.com/1/005/647.htm)｜[7 月首次亮相](https://finance.sina.com.cn/jjxw/2026-07-17/doc-iniickay1429452.shtml)

2. **“该不该打断”终于有了专门模型。** 9 月 18 日论文 *I'll Keep an Ear Out* 让 AudioLLM 在持续音频里输出“提醒/保持安静”，同时学习事件刚出现、持续相关、无关抑制和去重；作者报告流式平均延迟 3.5 秒。这很适合做耳机插播前的决策层，但论文面向可穿戴辅助场景，不等于手机低功耗常驻，也未在论文页发现可直接复现的代码入口，指标仍是作者结果。[论文](https://arxiv.org/abs/2609.21183)

3. **实时语音后端也补了两块。** 阿里云 9 月 20 日上线 Qwen-Audio-3.1-Realtime-Plus，官方列出全双工、函数调用与 Web 搜索，可做中文语音入口；Bolna 0.10.251 则在 9 月 22 日修复 Sarvam TTS 通话中 WebSocket 掉线后的接收恢复。它们能补“边聊边办事”和链路恢复，但都不负责耳机佩戴判断、消息确认或真实蜂窝电话线路。[Qwen](https://www.alibabacloud.com/help/en/model-studio/qwen-audio-3-1-realtime-plus)｜[Bolna](https://github.com/bolna-ai/bolna/releases/tag/0.10.251)

电话兜底复核无变化：Twilio 官方仍写明中国大陆呼入、呼出均不支持；香港和新加坡仍可双向，具体号码与来显条件各自受当地规则约束。[大陆](https://www.twilio.com/en-us/guidelines/cn/voice)｜[香港](https://www.twilio.com/en-us/guidelines/hk/voice)｜[新加坡](https://www.twilio.com/en-us/guidelines/sg/voice)

最值得试：先把“提醒/静默”决策器接到文字未读状态，再决定是否在确认佩戴时播到耳机；电话继续只做高优先级兜底。

来源：官网/仓库已查｜国内已查｜论文已查｜X公开检索未找到可核的新官方/作者帖，覆盖受限。
