中文

15 秒了解

家里已有的音箱继续用它们自己的麦克风和喇叭，但唤醒词、对话、记忆、音乐都走你自己的服务：一台机器上的 rui 服务 + 每种音箱一个桥接插件。控制台 /v2 可以加设备、改配置、看日志和轨迹，不用再手改 YAML。

说「你好小瑞」→ 它答「在呢」→ 说「放首歌」→ 播放中再说「下一首」→ 说「退下吧」→ 结束这一轮

架构

架构图源文件：docs/assets/architecture.mmd；导出物 architecture.svg / architecture.png（Mermaid；改配色后重新导出即可）。

三个仓库是什么关系

约束有两条，缺一不可：设备清单里的 compatibility.rui_min（这台设备至少要哪个 rui）与插件清单里的 rui_compat（这个插件支持哪一段 rui）。两者不满足时，加载器会拒绝这个插件，而不是带着半截功能跑。

LX06 的插件必须外置：把 rui-xiaoai-bridge 挂到 /opt/rui-plugins，并设 RUI_ADAPTER_PLUGINS + PYTHONPATH=/opt/rui-plugins；挂上后日志为 adapter plugins: registered dot2, lx06。核心不再有兼容别名、也不回退内置实现；不挂插件时设备错误会明确提示去 mount 桥。0.8.0 起旧配置里的 adapter: xiaoai 不再被迁移：读入即失败，报错含 devices[i] (room): adapter 'xiaoai' is no longer accepted in 0.8.0; change devices[].adapter to 'lx06'，把该行改成 lx06 即可。

让 AI 帮你装

把下面这段直接贴给任意 AI 助手（Claude / ChatGPT / DeepSeek 都行）：

目标：用 Docker 跑起 rui 0.7.9，接一台 Dot2 和一台 LX06。
前置：docker compose version 通过；8080 空闲；5 GB 磁盘。
1) git clone --branch main https://github.com/yes999zc/rui.git && cd rui/examples/two-speakers   # main 才有本示例；镜像仍 pin 在 0.7.9
2) git clone --branch v1.3.2 https://github.com/yes999zc/rui-dot2-bridge.git && git clone --branch v0.2.2 https://github.com/yes999zc/rui-xiaoai-bridge.git
3) mkdir -p plugins && cp -r rui-dot2-bridge/src/rui_dot2_bridge rui-xiaoai-bridge/src/rui_xiaoai_bridge plugins/
4) 写 ./config/config.yaml：web.auth_token、devices[]（id/adapter/room/host/listen_port；lx06 的 host 必填）
5) TLS（dot2 的 tls_port 开启时必需）：把原 controller 证书放到 ./secrets/（server.pem + server.key）；没有就运行 `bash ../../scripts/make_dev_tls.sh` 自签。自签只保证音频平面能起，真机是否接受取决于设备信任的 CA（详见 `docs/audit/tls-required-2026-10.md`）。**默认只起第一台设备的平面；两台都起需在 config 设 `multi_device: {enabled: true, max_planes: 2}`**
6) export RUI_DEVICE_TOKEN=<设备令牌>; docker compose up -d
7) 验收：/health 里 devices=2、device_errors=0、adapter_plugins.registered=["dot2","lx06"]；audio_plane=false 时看日志的 RUI_DEVICE_TOKEN / missing TLS material
失败看：docker compose logs --tail=80 rui 的 config error / adapter / TLS；回滚：down 后把 image 换回上一版 tag；TLS 失败可先移走 ./secrets/ 确认是否 TLS 引起
不需要自己编译 ffmpeg（镜像已内置）

特性

• 多设备、按房间隔离：设备按房间分组，唤醒与播放的互斥只在本房间生效，跨房间的播放与回合仲裁互不影响。
• 播放中打断（barge-in）：播放时唤醒立即停播并进入对话，服务端停播延迟以毫秒计。
• 自回声抑制：唤醒提示音（「在呢」）实测 0/10 泄漏；普通唤醒撞上进行中的告别语时，会丢弃已排队的我方语音。冷启动/重叠 <110 ms 的边界仍可能漏（见已知限制）。
• 可插拔设备：dot2 / xiaoai 两个桥接插件遵循同一份 rui-device/v1 契约，桥与核心可独立升级。
• 控制台 v2：设备、房间、配置分组、日志、轨迹、音乐、凭据、体检；中英双语，字段说明来自服务端目录。
• 控制台入口：只有 v2——/ 302 到 /v2/（保留 query）。0.8.0 删除了旧面板与 15 条旧链接（/panel/* 11 个面板 + /setup、/secrets、/diagnostics、/channels）的 302：/legacy 与这些路径现在都是 404；web.console 键仅为配置兼容保留，值被忽略。
• 本地优先：模型与密钥都在你自己的机器上；配置文件是唯一事实来源，控制台只提交你改过的字段。

三行启动

git clone --branch v0.7.9 https://github.com/yes999zc/rui.git && cd rui
RUI_IMAGE=ghcr.io/yes999zc/rui@sha256:5aa362f6b4f7d7fd24a7f01d41c31e1523ce2401296ac3b203b9f684a314b9f7 bash deploy/install.sh   # 0.7.9 amd64 摘要（默认 :latest）
cat console-token.txt                       # 控制台 token（也可用 rui token 重新读取）

装完打开 http://127.0.0.1:8080/v2/，用上面的 token 登录，在「设备」页添加第一台音箱。

外置设备插件（0.7.6 起）

核心镜像不再内置任何设备实现：一个核心镜像 + 一个插件目录即可，不需要造下游镜像。

1. 宿主目录：/vol1/rui/plugins/rui_dot2_bridge/、/vol1/rui/plugins/rui_xiaoai_bridge/（即两个桥仓的 src/<包名>）。
2. Compose 挂载 + 导入路径：-v ./plugins:/opt/rui-plugins:ro、PYTHONPATH=/opt/rui-plugins。
3. 显式注册：RUI_ADAPTER_PLUGINS=dot2=rui_dot2_bridge.adapter:build,lx06=rui_xiaoai_bridge.adapter:build。

直接可用的一份参考：examples/two-speakers/compose.yaml（两个桥 + TLS + 两段注册一次到位）；只想先跑单台核心可以走上面的「三行启动」（deploy/install.sh）。

挂载后日志出现 adapter plugins: registered dot2, lx06；/health.planes 逐台显示 running/error/not_started。
不挂插件时设备错误会直接提示去 mount 对应桥；旧配置的 adapter: xiaoai 在 0.8.0 读入即报错（给出改成 lx06 的修复句），不再有读入迁移与弃用告警。

兼容性与前置条件

• Python 3.11+；Docker 与 Docker Compose。推荐镜像（amd64 平台摘要，老版 docker 不认 OCI index）：
  ghcr.io/yes999zc/rui@sha256:5aa362f6b4f7d7fd24a7f01d41c31e1523ce2401296ac3b203b9f684a314b9f7（tag 0.7.9 指向同一构建）。
• 需要自备 LLM / TTS / ASR 供应商（在配置文件里填，或使用本地模型）。没有凭据时服务能起，但对话链路不会工作。
• 设备前置条件（第三方设备，与厂商无官方关联）：
  • EchoMuse Dot2（RS03QR）：固件需已具备设备本地接口，并与本服务在同一局域网；设备侧音频参数由桥接插件下发。
  • 小米 LX06 一类设备：仅支持已开放本地接口的固件版本；本项目不提供、也不建议任何刷机或越权步骤。
• 播放音乐需要自备音乐源（如 Navidrome）；未配置时「放首歌」会明确报错，不会静默失败。

已知限制：未声明 wake_words_residual 的第三方清单不做 vendor 短语兜底（vendor 广播/回声可能进入回合）；两音箱默认只起第一台设备的平面，需 multi_device: {enabled: true, max_planes: 2}；自签证书只保证音频平面能启动，真机是否接受取决于设备信任的 CA。（完整清单见下文「已知限制」一节与 docs/known-limitations.md。）

支持矩阵

音量刻度：控制台与语音用的是百分比；LX06 的设备刻度是 0–255，30% 对应 76（≈30×2.55），属正常映射，不是故障。dot2 的设备刻度是 0–127（0.5 dB/step），DOT2_VOLUME_CAP 是百分比上限（adapter 默认 100）。
| 指示灯 | ✅ 实测配色（2026-10-27 人眼）：唤醒=绿、思考=蓝、说话=青、放歌=紫；事件链 listening → thinking → speaking → off 与日志一致 | ✅ 同配色（2026-10-27 人眼）：唤醒=绿、思考=蓝、说话=青、放歌=紫 |
| 多设备并发 | 两台平面可同时 running；同时发起第二个对话会被 409 拒绝（a turn is already running，单会话 runner，2026-10-27） | 同左（同一 runner） |
| 音乐中短指令（下一首 / 停止播放） | 10/10（2026-10-26，人声，1 m） | 未测 |

安装路径实测（2026-10-26）

在 examples/two-speakers/ 下按本文件的步骤起容器（镜像 ghcr.io/yes999zc/rui@sha256:5aa362f6b4f7d7fd24a7f01d41c31e1523ce2401296ac3b203b9f684a314b9f7，两个插件目录 bind mount）：

$ curl -s localhost:8080/health | python3 -c 'import json,sys;d=json.load(sys.stdin);print({k:d[k] for k in ("status","version","devices","enabled_devices","device_errors","audio_plane","adapter_plugins")})'
{'status': 'ok', 'version': '0.7.9', 'devices': 2, 'enabled_devices': 2, 'device_errors': 0, 'audio_plane': False, 'adapter_plugins': {'registered': ['dot2', 'lx06'], 'warnings': 0}}

如实说明：adapter_plugins.registered、devices、device_errors 在任何情况下都成立。audio_plane 取决于「哪台设备的平面被启动」（实测见 docs/audit/tls-required-2026-10.md）：

• 默认 multi_device.enabled=false：单平面，只启动配置里的第一台设备。第一台是 dot2（listen_port 即 tls_port）且 secrets/server.pem|key 缺失 → audio_plane=false，日志 missing TLS material；有证书（原证书或自签）→ true，wss listening。第一台是 xiaoai 且 RUI_DEVICE_TOKEN 为空 → false，日志 RUI_DEVICE_TOKEN is empty。
• 两台都起：需 multi_device.enabled=true 且 max_planes >= 2，按台起平面、单台失败不拖垮其他；最小示例见 config.example.yaml 的 multi_device 注释，/health 的 planes 逐台显示状态。本卡实测的是默认单平面。
• 两个 token + 证书齐备、dot2 在前时实测 audio_plane=true。
  自签只解决「平面能起」，真机是否接受取决于设备信任的 CA；原 controller 证书丢失时设备需重新配网。

故障排查

已知限制

完整清单（编号 L1–L16）在 docs/known-limitations.md；本节只列对外必须说的。

• HE-02 回声切除仍有缺口：唤醒提示音的回声在「冷启动 + 不削波 + 设备播放延迟 > 800 ms」或「用户与提示音重叠 < 110 ms」时仍可能漏进 ASR；更细的门限实测是设备延迟 900–1200 ms 时不削波、1200 ms 削尾部、1200 ms 重叠时漏切（docs/known-limitations.md L1）。
• 0.7.9 的意图层修复：N1/N3 已在真机验过，N2 仍待按新口径复验：2026-10-28 08:54–09:33（本地；UTC 00:54–01:33）在生产真机上确认了 C1（结束词 + 有歌 = 停歌且关会话）、暂停/结束词语义与唤醒应答改序，docs/acceptance-pending.md 已勾 N1/N3/ACK-1/ACK-2。但仍不要把 0.7.9 描述成「意图层已修好」：N2 的判据被 d.6（b91599a）有意反转为「停止你好 命中结束词」，旧口径的真机记录作废，需按新口径复验 2 min；L17 唤醒回声修复（82d45d9，11:47）发生在那次真机之后，仍只经 headless 验证。
• 对话运行器目前是单会话：一台设备唤醒时会中止另一台正在合成的回复（按房间隔离目前只覆盖播放与回合仲裁）；第二台同时发起对话会被 409 拒绝（2026-10-27 实测，见支持矩阵）。
• 已播出的告别语尾音无法追回，只能阻止后续语音块。
• 唤醒词暂不按房间隔离。两台同型号 dot2 从未验证（现网与全部真机记录都是 dot2 + lx06 异型号），同屋双设备的行为见 docs/backlog-0.8.md §c。

致谢

致谢

这个项目站在很多人的工作上。按它们在这套系统里的实际位置分两类；许可证与分发细节见 THIRD_PARTY.md。

跑在链路里的

• @coderzc/open-xiaoai（MIT）—— LX06 的设备端客户端与 WebSocket 协议；设备靠它拨进来，核心只讲协议、不内置它的代码。
• k2-fsa/sherpa-onnx（Apache-2.0）—— 本地 ASR 与关键词唤醒运行时，镜像里的唤醒模型与 VAD 也经由它的 release 分发。
• snakers4/silero-vad（MIT）—— 语音活动检测模型。
• FunAudioLLM/SenseVoice / modelscope/FunASR —— 本地中文 ASR 模型；权重按 FunASR Model License 使用，rui fetch-models 下载，不随镜像分发。
• FFmpeg（LGPL-2.1-or-later，构建时禁用 GPL/nonfree）—— 音乐转码与渠道语音编码；镜像里随附 libopus、libmp3lame 的许可证。
• xiph/speexdsp（BSD-3-Clause）—— 回声消除，播放中打断靠它。
• openWakeWord（Apache-2.0）—— 设备端 OWW 模型（如 hey_jarvis）来自这套方案；目前设备端唤醒是关闭的，走服务端 KWS。

设计上参考了公开做法

• Rhasspy 的分域式服务拆分、rhasspy/wyoming 的协议层抽象。
• wilbowes/EchoMuse —— 设备管理控制台的状态机与「可解释向导」思路（未复制代码）。
• openWakeWord 对「未测 / 限制」照实标注的习惯，我们也沿用了。

音乐播放对接 Navidrome / OpenSubsonic；控制台用 Vue + Vite，服务端 FastAPI。实现代码除上述依赖外均为本仓库原创。

License

MIT。

English

In 15 seconds

Your existing speakers keep their own microphones and speakers, but wake words, dialogue, memory and music run on your own machine: one rui service plus one bridge plug per speaker model. The /v2 console adds devices, edits configuration and shows logs and traces — no YAML editing required.

"Hey Rui" → "I'm here" → "play some music" → "next" (while playing) → "stand down" → turn ends

Architecture

Source: docs/assets/architecture.mmd; exports architecture.svg / architecture.png (Mermaid; re-export after changing colours).

How the three repositories fit together

Two constraints must both hold: compatibility.rui_min in the device manifest (the minimum core this device needs) and rui_compat in the plug manifest (the core range the plug supports). If either fails, the loader refuses the plug instead of running it half-working.

The LX06 plug is external only: mount rui-xiaoai-bridge under /opt/rui-plugins and set RUI_ADAPTER_PLUGINS + PYTHONPATH=/opt/rui-plugins; with it mounted the log reads adapter plugins: registered dot2, lx06. There is no compatibility alias and no built-in fallback, and without the plug the device error tells you to mount the bridge. Since 0.8.0 a legacy adapter: xiaoai is no longer migrated: it fails on load with devices[i] (room): adapter 'xiaoai' is no longer accepted in 0.8.0; change devices[].adapter to 'lx06'. Change that line to lx06; do not wait for a fallback.

Let an AI install it

Paste this into any assistant (Claude / ChatGPT / DeepSeek):

Goal: run rui 0.7.9 with Docker, one Dot2 and one LX06.
Preflight: docker compose version; port 8080 free; 5 GB disk.
1) git clone --branch main https://github.com/yes999zc/rui.git && cd rui/examples/two-speakers   # main 才有本示例；镜像仍 pin 在 0.7.9
2) git clone --branch v1.3.2 https://github.com/yes999zc/rui-dot2-bridge.git && git clone --branch v0.2.2 https://github.com/yes999zc/rui-xiaoai-bridge.git
3) mkdir -p plugins && cp -r rui-dot2-bridge/src/rui_dot2_bridge rui-xiaoai-bridge/src/rui_xiaoai_bridge plugins/
4) write ./config/config.yaml: web.auth_token and devices[] (id/adapter/room/host/listen_port; host required for lx06)
5) TLS (required when dot2's tls_port is on): put the original controller cert into ./secrets/ (server.pem + server.key); if you do not have it, run `bash ../../scripts/make_dev_tls.sh`. Self-signed only proves the audio plane starts; whether the device accepts it depends on its trusted CA (see `docs/audit/tls-required-2026-10.md`). **By default only the first configured device gets a plane; to run both, set `multi_device: {enabled: true, max_planes: 2}` in the config**
6) export RUI_DEVICE_TOKEN=<device token>; docker compose up -d
7) verify: /health shows devices=2, device_errors=0, adapter_plugins.registered=["dot2","lx06"]; if audio_plane=false read the log for RUI_DEVICE_TOKEN / missing TLS material
On failure: docker compose logs --tail=80 rui (config error / adapter / TLS); rollback: down, switch image back; for TLS failures move ./secrets/ away first to isolate
No ffmpeg build needed (the image ships it)

Features

• Per-room isolation: devices are grouped by room; wake, playback and turn arbitration are scoped to a room (the dialogue runner itself is still single-session).
• Barge-in: a wake during playback stops the audio immediately and starts a turn, with millisecond-scale stop latency on the server side.
• Self-echo suppression: the acknowledgement cue leaked 0/10 on the device; a stale goodbye still queued is dropped when a wake arrives. Cold-start or <110 ms overlap can still leak (see Known limitations).
• Pluggable devices: the dot2 and xiaoai bridges share one rui-device/v1 contract and version independently.
• Console v2: devices, rooms, grouped configuration, logs, traces, music, credentials and health checks; bilingual UI fed by a server-side catalog.
• Console entry: v2 only — / answers a 302 to /v2/ (query preserved). 0.8.0 removed the legacy panel shell and the 15 old links (the 11 /panel/* panels plus /setup, /secrets, /diagnostics, /channels): /legacy and those paths are 404 now. web.console stays for config compatibility only; its value is ignored.
• Local first: models and keys stay on your machine; the config file is the source of truth and the console submits only changed fields.

Three-line quickstart

git clone --branch v0.7.9 https://github.com/yes999zc/rui.git && cd rui
RUI_IMAGE=ghcr.io/yes999zc/rui@sha256:5aa362f6b4f7d7fd24a7f01d41c31e1523ce2401296ac3b203b9f684a314b9f7 bash deploy/install.sh   # 0.7.9 amd64 digest (default: :latest)
cat console-token.txt                       # console token (or: rui token)

Open http://127.0.0.1:8080/v2/, log in with that token, and add your first speaker on the Devices page.

External device plugs (since 0.7.6)

The core image ships no device implementation: one core image plus a plug directory is enough, and no downstream image is needed.

1. Host directory: /vol1/rui/plugins/rui_dot2_bridge/, /vol1/rui/plugins/rui_xiaoai_bridge/ (the src/<package> of each bridge repo).
2. Compose mount + import path: -v ./plugins:/opt/rui-plugins:ro, PYTHONPATH=/opt/rui-plugins.
3. Explicit registration: RUI_ADAPTER_PLUGINS=dot2=rui_dot2_bridge.adapter:build,lx06=rui_xiaoai_bridge.adapter:build.

One ready-to-copy reference: examples/two-speakers/compose.yaml (both plugs, TLS and registration in one file). For a core-only first run use the three-line install above (deploy/install.sh).

After mounting, the log says adapter plugins: registered dot2, lx06; /health.planes shows each device as running/error/not_started.
Without a plug the device error tells you which bridge to mount; since 0.8.0 a legacy adapter: xiaoai fails on load with the fix sentence naming lx06 (no read-time migration, no warning-only path).

Compatibility and prerequisites

• Python 3.11+, Docker and Docker Compose. Recommended image (amd64 platform digest; old docker cannot resolve the OCI index):
  ghcr.io/yes999zc/rui@sha256:5aa362f6b4f7d7fd24a7f01d41c31e1523ce2401296ac3b203b9f684a314b9f7 (tag 0.7.9 points at the same build).
• Your own LLM / TTS / ASR providers (configured in the config file, or local models). The service starts without credentials, but the dialogue path stays idle.
• Device prerequisites (third-party hardware, not affiliated with any vendor):
  • EchoMuse Dot2 (RS03QR): firmware must already expose the device's local interface and be on the same LAN; audio parameters are pushed by the bridge.
  • Xiaomi LX06-class devices: only firmware that already exposes a local interface is supported. No flashing, rooting or privilege-escalation steps are provided or endorsed.
• Music needs your own source (e.g. Navidrome); without one, "play some music" fails loudly instead of silently.

Known limitations: a third-party manifest that does not declare wake_words_residual gets no vendor-phrase backstop (vendor broadcasts/echoes may enter a turn); two speakers only start one plane by default unless multi_device: {enabled: true, max_planes: 2}; a self-signed certificate only lets the audio plane start — real-device acceptance depends on the CA the device trusts. (The full list is the Known limitations section below and docs/known-limitations.md.)

Support matrix

Volume scale: console and voice use percent; the LX06 device scale is 0–255, so 30% is 76 (≈30×2.55) — expected, not a fault. dot2's device scale is 0–127 (0.5 dB/step); DOT2_VOLUME_CAP is a percent ceiling (adapter default 100).
| LED | ✅ observed (2026-10-27, human eye): wake=green, thinking=blue, speaking=cyan, music=purple; matches the listening → thinking → speaking → off log chain | ✅ same colours (2026-10-27, human eye): wake=green, thinking=blue, speaking=cyan, music=purple |
| Concurrent multi-device | both planes run at once, but a second simultaneous turn is refused with 409 (a turn is already running, single-session runner; 2026-10-27) | same (one runner) |
| Short commands during music (next / stop) | 10/10 (2026-10-26, human voice, 1 m) | untested |

Install-path proof (2026-10-26)

Started from examples/two-speakers/ following the steps in this file (image ghcr.io/yes999zc/rui@sha256:5aa362f6b4f7d7fd24a7f01d41c31e1523ce2401296ac3b203b9f684a314b9f7, both plug dirs bind-mounted):

$ curl -s localhost:8080/health | python3 -c 'import json,sys;d=json.load(sys.stdin);print({k:d[k] for k in ("status","version","devices","enabled_devices","device_errors","audio_plane","adapter_plugins")})'
{'status': 'ok', 'version': '0.7.9', 'devices': 2, 'enabled_devices': 2, 'device_errors': 0, 'audio_plane': False, 'adapter_plugins': {'registered': ['dot2', 'lx06'], 'warnings': 0}}

To be precise: adapter_plugins.registered, devices and device_errors always hold. audio_plane depends on which device's plane is started (measured; see docs/audit/tls-required-2026-10.md):

• Default multi_device.enabled=false: one plane, only the first configured device. dot2 first (listen_port is tls_port) without secrets/server.pem|key → audio_plane=false, log missing TLS material; with a certificate (original or self-signed) → true, wss listening. xiaoai first with an empty RUI_DEVICE_TOKEN → false, log RUI_DEVICE_TOKEN is empty.
• Both devices: set multi_device.enabled=true and max_planes >= 2; each device gets a plane and one failure does not take down the others. See the commented multi_device block in config.example.yaml; /health.planes shows each device's state. This card measured the default single-plane path.
• With both tokens and the certificate, dot2 first, the measured value is audio_plane=true.
  Self-signed only makes the plane start; real-device acceptance depends on the device's trusted CA, and a lost original controller certificate means the device must be re-provisioned.

Troubleshooting

Known limitations

The full list (entries L1–L16) lives in docs/known-limitations.md; this section only carries what users must know.

• The HE-02 echo-excision gap is still open: the acknowledgement cue can still slip into ASR on a cold start with no clipping and >800 ms device latency, or when the user overlaps the cue by less than 110 ms; the finer thresholds are: no excision at 900–1200 ms device lag, tail clipped at 1200 ms, missed when overlapped at 1200 ms (docs/known-limitations.md L1).
• The 0.7.9 intent-layer fix: N1/N3 are verified on a real device, but N2 still needs a re-check: production hardware on 2026-10-28 (08:54–09:33 local, 00:54–01:33 UTC) confirmed C1 (an end word with music playing stops the song and closes the session), the pause/end-word semantics and the reordered wake acknowledgement, and docs/acceptance-pending.md ticks N1/N3/ACK-1/ACK-2. Do not describe 0.7.9 as "the intent layer is fixed" yet: b91599a (d.6) deliberately inverted N2 to "停止你好 matches the end word", which voids the old-criterion hardware record, so N2 needs a 2 min re-check under the new wording; the L17 ack-echo fix (82d45d9, 11:47) landed after that run and is still headless-only.
• The dialogue runner is single-session: a wake on one device aborts a reply being synthesised for another (per-room isolation currently covers playback and turn arbitration only); a second simultaneous turn is refused with 409 (measured 2026-10-27, see the support matrix).
• A goodbye that has already been spoken cannot be recalled; only queued chunks are dropped.
• Wake words are not yet isolated per room. Two same-model Dot2 units have never been verified (production and every real-device record use the mixed dot2 + lx06 pair); see docs/backlog-0.8.md §c for the two-in-one-room behaviour we still have to measure.

Credits

Acknowledgements

This project stands on other people's work. Grouped by where each one actually sits; licenses and distribution details live in THIRD_PARTY.md.

In the pipeline

• @coderzc/open-xiaoai (MIT) — the LX06 device client and WebSocket protocol. The speaker dials in through it; we speak the protocol and ship none of its code.
• k2-fsa/sherpa-onnx (Apache-2.0) — local ASR and keyword-spotting runtime; the wake-word and VAD models in the image come through its releases.
• snakers4/silero-vad (MIT) — voice-activity detection model.
• FunAudioLLM/SenseVoice / modelscope/FunASR — the local Chinese ASR model; weights are used under the FunASR Model License and fetched at runtime, not shipped.
• FFmpeg (LGPL-2.1-or-later, built without GPL/nonfree) — music transcoding and channel voice encoding; libopus and libmp3lame notices ship with the image.
• xiph/speexdsp (BSD-3-Clause) — echo cancellation, which is what makes barge-in work.
• openWakeWord (Apache-2.0) — the device-side OWW models (e.g. hey_jarvis); on-device wake is off today and the server-side KWS does the work.

Borrowed in design

• Rhasspy's domain split and rhasspy/wyoming's protocol layering.
• wilbowes/EchoMuse — state machine and explainable-wizard ideas for the device console (no code copied).
• openWakeWord's habit of labelling untested areas plainly.

Music talks to Navidrome / OpenSubsonic; the console is Vue + Vite on top of FastAPI. Everything else here is original to this repository.

License

MIT.