语音桥是一台低成本语音设备，服务对象是不在学校体系内的失学儿童——平板、识字 App、课堂软件都触达不到他们。我们把整个过程公开做出来，并且正在寻找愿意和我们一起落地的在地机构。 VoiceBridge is a low-cost voice device for children who are out of school — the ones classroom tablets and literacy apps never reach. We are building it in the open, and we are looking for the local organisations who will run it with us.
在非洲，太阳能平板 + 本地语言识字 App + 广播电视教育已经被做出来了，有随机对照试验的证据，正在交给政府接管。那些项目服务的是在校儿童。我们不在这个赛道上竞争。 In Africa, solar tablets, local-language literacy apps and broadcast lessons already exist — with randomised-trial evidence and governments taking them over. Those programmes serve children who are in school. That is not our lane.
太阳能平板、本地语言课程、电视与广播教学。有证据、有规模、有政府在接管。我们做不出更好的，也不打算试。 Solar tablets, mother-tongue courseware, TV and radio lessons. Evidence-backed, scaling, and being handed to ministries. We cannot do it better and we are not going to try.
没有老师、没有基础，家里可能没人识字，很多孩子还要干活或照看弟妹。而一个识字 App 默认了孩子在学校、有老师、已具备阅读起点——三条他都不满足。 No teacher, no baseline, often no literacy at home, and many are working or caring for siblings. A literacy app quietly assumes school, a teacher, and a starting level of reading. None of the three holds.
一个不识字的孩子，依然能听、能说。按键说话把"会阅读"这个前置条件整个拿掉，而且它是最便宜的交互方式——一个按钮、一个麦克风、一个喇叭。 A child who cannot read can still listen and speak. Push-to-talk removes the literacy prerequisite entirely, and it is the cheapest interface there is: one button, one microphone, one speaker.
但语音也是风险最大的一环。语音识别在儿童语音和低资源语言上的错误率，远高于成人——某些低资源场景下词错误率可达 30%–80%。我们会在任何试点之前公布实测数据；如果它不成立，我们就公开说明，然后改设计。有一点我们不碰运气：识别可以只求"听个大概"，但孩子听到的内容不能——它是预先审核、预先录制的，不是实时生成的。 Voice is also the riskiest part. Speech recognition on children's speech in low-resource languages performs far worse than on adults — word error rates of 30–80% have been reported in some low-resource settings. We will publish our own measured error rate before any pilot, and if the approach fails we will say so and change the design. One thing we do not leave to chance: recognition is allowed to be approximate, but what the child hears is not — it is pre-vetted and pre-recorded, not generated live.
设备采集语音并送出去。音频不落盘。 The device captures speech and streams it. Audio is never written to disk.
识别 → 匹配预审回答库 → 播放预录的真人语音。最常用的回答由母语配音员预先录好、逐条人工审核，长尾才用合成语音。置信度不够时不猜，请孩子再说一遍。断网时改播本地缓存课程。 Transcribe → match against a pre-vetted answer library → play a recorded human voice. The most common replies are pre-recorded by native speakers and reviewed one by one; only the long tail is synthesised. When confidence is low it does not guess and asks the child to repeat. With no network it falls back to cached lessons.
越界或不当请求立即中止，转交现场协调员。 Off-topic or unsafe requests are stopped and handed to the coordinator.
孩子轮流用传给他的近场麦克风。听不懂、答不上、内容越界，一律转给协调员——他在场是使用前提，也是唯一的人工防线。 Children take turns with the pass-around near-field microphone. Anything unclear, unanswered or out of bounds goes to the coordinator — whose presence is a precondition for use, and who is the only human backstop.
通话两端没有真人。设备不接远程志愿者或老师，也不做设备间通话——成本和儿童保护风险因此大幅下降，代价是纠错只能靠现场协调员。但孩子听到的内容不是模型即兴生成的：常用回答由母语者预先录音、逐条人工审核，因为"说得必须准"这条承诺，不能用生成式模型来兑现。 No human is on the call. No remote volunteers or teachers, no device-to-device calling — cost and safeguarding risk both drop sharply, and the price is that correcting a wrong answer rests entirely on the coordinator in the room. But what the child hears is not generated on the fly: common answers are pre-recorded by native speakers and reviewed one by one, because a promise that the output must be correct is not one a generative model can keep.
下面这个模拟器可以真的按住说话。它跑的是预置脚本，不是实时 AI——但流程、拦截逻辑和离线降级都是按真实设计来的。 The simulator below actually responds to press-and-hold. It runs on a fixed script, not live AI — but the flow, the blocking logic and the offline fallback are the real design.
这是演示，不是实时 AI。回答来自预置脚本，用于展示交互流程、内容拦截和离线降级。真实识别率取决于语言与环境——我们会在试点前公开实测的词错误率，不会拿演示效果对外宣称。有一点演示是如实反映的：正式系统里，回答也不是即兴生成的，而是从预审回答库中匹配、由母语者预录的真人语音播出。演示按中心式场景设计：孩子拿到近场麦克风再说话。 This is a demo, not live AI. Replies come from a fixed script, used to show the interaction flow, content blocking and offline fallback. Real accuracy depends on language and environment — we will publish measured word error rates before any pilot and will not claim demo performance as real. One thing the demo does reflect faithfully: in production the reply is not generated on the fly either — it is matched from a pre-vetted library and voiced by a pre-recorded native speaker. The demo follows the shared-hub scenario: the child speaks into a passed-around near-field microphone.
路线：先用 1–2 台中心式验证。有人用吗、内容对吗、联网稳吗——这三个问题不需要 50 台就能回答。验证通过之后再铺手持式。 Route: validate with one or two hubs first. Do children use it, is the content right, is the connection stable — none of those needs 50 units. Handhelds follow only once that passes.
会话被计数，不被记录。后台只有设备编号、时间、时长和话题类别。 Sessions are counted, not recorded. The dashboard holds device ID, time, duration and topic category only.
我们是一个没有注册的志愿者团队，没有资金，也没有当地落地能力。我们有的只是一套公开的技术方案和愿意干活的人。所以这件事只有一种成立方式：在地机构主导。 We are an unregistered volunteer team with no funding and no ability to operate in the field. What we have is an open technical design and people willing to work. So there is exactly one way this works: a local organisation leads it.
你提供：场地、语言、儿童保护体系、持续运营、现场协调员。
我们提供：设备（先中心式，验证通过后再加手持）、技术维护、设备管理后台、全部文档开源。
You bring: the site, the language, safeguarding, continuity, and the coordinator.
We bring: the hardware — hubs first, handhelds after validation — technical upkeep, device dashboard, fully open documentation.
你已经让孩子稳定出现在同一个屋子里——这恰恰是最难的一步，也是我们最缺的东西。设备挂靠在你的活动上，而不是另起一个新项目。 You already get children into the same room regularly — that is the hardest part and the thing we lack most. The device attaches to your existing sessions rather than starting a new programme.
母语内容、课程对齐、本地化审校。已有成熟离线内容平台的机构，我们更愿意直接用你们的，而不是自己造。 Mother-tongue content, curriculum alignment, localisation review. If an offline content platform already exists, we would rather use yours than build another.
我们自己不接收捐款。所有资金由合作方或具备资质的托管机构收取、支出和公示。 We do not accept donations ourselves. All funds are received, spent and reported by a partner or a qualified fiscal sponsor.
实地 / 驻国志愿者。在合作方和地点确定之前，我们一个都不招。宁可晚一点，也不要承诺一个兑现不了的岗位——那对你是浪费时间，对当地的孩子更不负责任。 Field and in-country roles. We recruit none until a local partner and a site are confirmed. Better late than promising a placement we cannot deliver — that wastes your time and is worse for the children involved.
录 20 个目标孩子的真实语音，跑一遍候选引擎，把词错误率原样公布。Record real speech from 20 target children, run it through candidate engines, publish the word error rate as measured.
噪声、电源、断网、孩子真的会用吗。再加一项：同一个孩子近场与远场各录一遍，把识别率差距量化出来。Noise, power, dropouts, whether children actually use it. Plus one more: record the same child near-field and far-field, and quantify the accuracy gap.
与在地合作方共同执行，月度公开使用与支出摘要。验证通过后再铺手持式。Run jointly with a local partner, with monthly public usage and spending summaries. Handhelds follow only if this passes.
达不到门槛就停下，并把停的原因写出来。If thresholds are missed, we stop — and we publish why.
门槛已写进公开文档Thresholds are published我们最需要的是被反驳。告诉我们这个设想哪里在你们那里行不通——这比一封支持信有用得多。 What we need most is to be contradicted. Tell us where this design breaks in your context — that is worth far more to us than a letter of support.