E[E]C[C]H[H]O[O]M[M]I[I]N[N]D[D]

无论是参加一场工作面试，向领导汇报项目进展，还是向别人介绍一个复杂的想法，口头表达几乎贯穿了我们的工作与生活。但它有一个特点：当我们专注于思考下一句话该说什么时，往往很难同时观察自己正在怎样表达。Job interviews, progress reports, explaining a complicated idea — spoken expression runs through work and life. Yet while you focus on the next sentence, you can hardly watch how you are speaking at the same time.

有时候明明已经在脑海中组织好了内容，真正说出来时却变得零散；可能会频繁使用「嗯」「然后」「就是」，也可能反复解释同一个观点，或者因为不知道下一句该说什么而出现长时间停顿。A thought that felt organised comes out scattered. Fillers such as 嗯, 然后 or 就是 creep in, one point gets explained twice, and a long pause appears when the next sentence will not come.

还有一些问题更加隐蔽：以为自己回答了面试官的问题，实际上没有提供具体的案例和证据；以为完成了一次工作汇报，听众却仍不知道项目取得了什么进展、遇到了什么风险，以及需要他们做出什么决定。于是有了最初的问题——能不能做一个工具，把说过的话记录下来，再帮助我们理解自己的表达问题？Other problems hide better: you believe the question was answered, yet no concrete example or evidence was given; you believe the report is done, while the audience still does not know the progress, the risks, or what decision is needed. Hence the first question: could a tool record what was said, and help us understand our own expression?

如果只统计语气词、重复和停顿，产品很容易变成一个表达检测器：它可以说出你讲了多少次「嗯」、停顿了几次，却回答不了更重要的问题——这段表达是否完成了它原本的目的？面试回答要让面试官理解候选人的能力和经历；工作汇报要帮助听众了解状态、判断问题，或者做出决策。两者同属口头表达，目的、结构和评价标准却并不相同。Counting fillers, repeats and pauses alone turns the product into a detector: it can report how many times you said 嗯, but not whether the answer achieved its purpose. An interview answer must convey capability and experience; a report must help the audience understand status, judge a problem or make a decision. Same medium, different goals, structures and standards.

所以我没有把产品做成不区分场景的通用打分工具，而是先选择两个目标明确、又有实际使用需求的场景。So instead of a scene-agnostic scoring tool, the first version picked two scenes with clear goals and real demand.

对于模拟面试，我没有采用完全随机的聊天方式。开始练习前，用户需要填写目标岗位和岗位描述（JD），让 AI 理解岗位关注的能力，并据此生成更有针对性的问题。专业问题数量可以设置为 1 至 15 道，还可以选择是否加入自我介绍与最后的反问面试官环节。Interview practice is not a free-form chat. Before starting, the user fills in the target role and job description so the AI understands which capabilities matter and generates pointed questions. Role questions can be set from 1 to 15, with optional self-introduction and candidate-question stages.

语音识别是 EchoMind 的核心基础能力之一。练习过程中，用户通过麦克风输入语音，系统在浏览器内部进行识别并持续更新转写内容——不必等到录音结束才知道自己说了什么，文字会随表达过程逐步出现，让人可以观察正在形成的内容。Speech recognition is a core capability here. While practising, the browser recognises microphone input and keeps updating the transcript, so you can watch the words take shape instead of waiting until the recording ends.

开发过程中有一个会直接影响评分公平性的问题：如果语音识别把某个专业术语听错了，AI 是否应该因此降低用户的表达评价？例如用户在介绍软件项目时用了某个英文技术名称，模型却识别成了别的词——把这段错误的转写直接交给 AI，评价就会受到无关因素影响。因此文字记录被拆成两种形式。One question directly affects fairness: if recognition mishears a technical term, should the AI score the answer lower? A product name mis-heard as another word would drag the evaluation down for no reason. So the text record is split in two.

大模型很容易生成听起来合理、却没有指出问题所在的评价：说「你的回答缺少具体证据」，却不说是哪一部分不够具体；认为结构完整，却指不出实际支持判断的内容。因此，对于需要评分的维度，AI 必须提供来自语义稿的真实文字证据，评价结果不能凭空引用用户没有说过的话。系统会检查模型返回的证据是否真实存在于当前语义稿中——找不到时，该维度不能保留一个看似确定的分数，而应显示证据不足或无法评价。Models readily produce plausible feedback that never locates the problem: “your answer lacks concrete evidence” without saying which part, or “the structure works” without naming what supports it. So every scored dimension must quote real text from the semantic transcript; nothing may be attributed to the user that they did not say. The system verifies each quote against that transcript — when it is missing, the dimension shows insufficient evidence or unable to evaluate rather than a confident number.

侧边栏提供练习历史，按练习时间排列，可以打开已完成的报告，也可以恢复尚未完成的练习。历史记录保留当前场景、题目、回答、语义稿和评价结果，因此不必每次都从头配置——一场模拟面试可能包含十几道题，用户可以在不同时间分阶段完成。The sidebar lists sessions by time: finished reports reopen, unfinished sessions resume. Scene, questions, answers, semantic transcripts and evaluations are kept, so nothing has to be configured twice — a fifteen-question interview can be completed in stages.

首版没有引入复杂的训练任务管理系统，而是优先保证练习历史能够被保存、恢复和查看。目前这些记录保存在当前浏览器的本地存储中，清除浏览器数据或更换设备时可能无法保留；跨设备同步与更完整的长期训练记录仍待探索。Version one skips a heavyweight training-task manager and instead guarantees that history can be saved, resumed and read. Records currently live in this browser's local storage, so clearing site data or switching devices may lose them; sync and durable long-term records remain open questions.

在界面设计中，我希望用户始终清楚三件事：我现在正在进行哪一项练习？我的声音是否被正常记录？接下来应该执行什么操作？因此 EchoMind 采用相对明确的页面结构。The interface should always answer three questions: which task am I in, is my voice being captured, and what do I do next? The page structure follows from that.

EchoMind 也是一次把语音 AI、前端交互与产品评价机制结合起来的开发实践。前端使用 React、TypeScript、Vite 和 Tailwind CSS，组件基于 shadcn、Radix UI 等工具组织，并结合 DaisyUI 实现多主题能力；语音识别使用 sherpa-onnx 相关技术，把本地中文语音模型运行在浏览器环境中；AI 内容分析通过兼容的模型 API 完成，并使用独立的请求代理处理模型连接。EchoMind is also a build practice that joins speech AI, front-end interaction and an evaluation model. The front end uses React, TypeScript, Vite and Tailwind CSS with components organised through shadcn and Radix UI, plus DaisyUI for theming. Recognition runs a local Chinese speech model in the browser via sherpa-onnx, and AI analysis goes through a compatible model API behind a dedicated request proxy.

一个表达训练产品，需要控制好记录、分析和用户干预之间的关系。随着功能逐步实现，这些原则逐渐明确，也构成了 EchoMind 当前的交互与评价基础。A training product has to balance recording, analysis and user intervention. As features landed, these principles settled — and they now underpin how EchoMind behaves.

最初，我只是希望能记录自己说过的话，并借助 AI 发现口头表达中的语气词、重复、停顿和逻辑问题。但随着产品逐渐展开，我发现「表达能力」并不是一个适合用单一分数衡量的对象：一段面试回答是否有效，取决于它是否切题、能否提供证据、是否说明了候选人的能力；一次工作汇报是否有效，则要看结论是否明确、事实是否充分、风险是否被说明，以及听众是否知道接下来该做什么。于是产品从通用表达检测，调整为围绕具体场景进行练习与评价的工具。It began as a way to record what I said and let AI surface fillers, repeats, pauses and logic problems. As the product grew, “expression ability” turned out to resist a single score: an interview answer works when it is on topic, evidenced and shows capability; a report works when the conclusion is clear, the facts sufficient, the risks stated and the audience knows what comes next. So the product moved from generic detection to scene-specific practice and evaluation.

交互上需要考虑用户如何进入练习、如何控制录音、什么时候确认回答、怎样回看证据；技术上要处理实时语音识别、本地模型运行、异步任务、文本修订、数据存储和性能问题；AI 评价上则要守住一条更重要的边界——模型可以提供分析，但分析结果必须尽可能有据可查，不能因为模型能生成一个分数，就默认这个分数可信。Interaction had to answer how a session starts, how recording is controlled, when an answer is confirmed and how evidence is revisited. Engineering had to handle live recognition, local models, async work, text revision, storage and performance. Evaluation had to hold one boundary above all: a model can analyse, but the analysis must stay traceable — a generated number is not automatically a true one.

EchoMind 目前仍处于 Beta 阶段，已经拥有基本的表达训练流程，但语音识别质量、长时间稳定性、模型评价一致性和长期训练效果仍需要更多测试与验证。最后留下一个最值得继续探索的问题：如何让 AI 在不替代用户思考和表达的前提下，帮助用户更准确地认识自己的表达方式，并找到可以实际执行的改进方向？EchoMind is still Beta. The basic training flow works, yet recognition quality, long-session stability, evaluation consistency and long-term outcomes all need more testing. One question is worth keeping: how can AI help people understand their own expression more accurately — and find changes they can actually make — without thinking or speaking for them?