---
{"schema":1,"date":"2026-09-21","timezone":"Asia/Shanghai","title":"电话语音 Agent 每日观察","revision":1,"publication":"publish","review_status":"generated","summary":["OpenAfon把唤醒、常驻脑、跨设备记忆和主动触达放进同一框架。","华为智慧播报补证：播放中只能同时播或不播，未见自动压音。","新论文把全双工语音前端接到可持久化的工具Agent后端。"],"events":[{"project":"openafon","title":"七日补漏：OpenAfon把唤醒、常驻脑、跨设备记忆、主动播报/推送与耳机插话放进同一自托管框架","kind":"code","event_date":null,"publication_date":null,"discovered_date":"2026-09-21","evidence":"verified","source":"https://github.com/iamvazghen/OpenWatari"},{"project":"huawei-smart-broadcast","title":"官方文档补证：音乐或视频播放期间，智慧播报仅提供“同时播报”或“不播报”，未声明自动压低或暂停第三方音频","kind":"other","event_date":null,"publication_date":null,"discovered_date":"2026-09-21","evidence":"source-claim","source":"https://consumer.huawei.com/cn/support/content/zh-cn16050593/"},{"project":"frontend-backend-duplex-toolcalls","title":"论文用全双工语音前端把工具任务委派给LangGraph后端，并用线程检查点跨轮保存会话状态","kind":"paper","event_date":"2026-09-16","publication_date":"2026-09-16","discovered_date":"2026-09-21","evidence":"source-claim","source":"https://arxiv.org/abs/2609.19334"},{"project":"bolna-0-10-245","title":"Bolna 0.10.245加入AssemblyAI universal-3.x流式模型与ASR关键词/上下文偏置","kind":"release","event_date":"2026-09-20","publication_date":"2026-09-20","discovered_date":"2026-09-21","evidence":"verified","source":"https://github.com/bolna-ai/bolna/releases/tag/0.10.245"},{"project":"aion-os-audio-focus","title":"旧项目补录：Android工程实现语音助手音频焦点、音乐降至30%、Bluetooth SCO与拔耳机处理；未提供真实入耳确认","kind":"code","event_date":null,"publication_date":null,"discovered_date":"2026-09-21","evidence":"verified","source":"https://github.com/Manthan-13521/Jarvis-AI-Assistant/blob/main/android/app/src/main/java/com/aion/os/voice/audio/AudioFocusManager.kt"}]}
---
# 电话语音 Agent 日报｜2026-09-21

今天没有出现把“耳机插播＋后台Agent＋电话兜底”一次做完的新成品，但补到了三块很接近目标的拼图。

1. **OpenAfon 更像目标架构骨架。** 七日补漏，不是今日发布。它把本地唤醒、流式语音、耳机插话、六层记忆、后台调度和主动提醒接到同一个常驻“脑”上，手机和电脑共享状态；没在听时可退到推送。限制是手机主要靠 Siri 快捷指令、Tasker/Termux，`phone-headphones` 只是客户端声明/连接状态，不是真正入耳检测；也未核到跨第三方播放器自动压音恢复或 PSTN 来电兜底。[仓库](https://github.com/iamvazghen/OpenWatari)｜[代码](https://github.com/iamvazghen/OpenWatari/blob/master/src/afon/edge/device_profile.py)

2. **华为智慧播报把跨 App 边界说清了。** 这是昨天 FreeBuds 7 的补证：HarmonyOS 6+ 可播通知栏消息，支持关键词、免打扰和部分场景的手势/头势确认；FreeBuds 7 还能追问刚播内容。关键是音乐/视频播放时，官方只提供“同时播报”或“不播报”，未声明自动 duck、暂停再恢复第三方媒体，所以仍不是无缝播放器式插播。[官方说明](https://consumer.huawei.com/cn/support/content/zh-cn16050593/)

3. **9 月 16 日论文把实时语音和后台办事拆开。** 全双工语音前端负责自然插话，需要工具时把流式转写交给 LangGraph 后端，后端用线程检查点跨轮保存状态，再把结果回灌语音。作者报告工具触发召回 92–97%、无关请求拒绝 81.2%；这是作者评测，且不是手机/耳机产品。在线演示已公开，本次未核到完整训练代码或权重。[论文](https://arxiv.org/abs/2609.19334)｜[演示](https://huggingface.co/spaces/frontend-backend-duplex/demo)

补充：Bolna 0.10.245 昨日加入流式识别关键词/上下文偏置。电话兜底复核仍是 Twilio 不支持外呼中国大陆；香港、新加坡可双向，但有本地号码/来显合规要求。[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：OpenAfon 常驻脑＋Android Audio Focus＋真实入耳状态，先跑通“确认佩戴→压音插播→明确确认→恢复播放”。

来源：官网/仓库已查｜国内已查｜论文已查｜X公开检索未获可核原帖，覆盖受限。