我不是程序员，用声音指挥 AI，独自做了一个有 20 万行代码的产品 I'm Not a Programmer. I Built a 200,000-Line Product Solo Using Voice-Controlled AI.

产品是什么 What is HeraldVox

唤达（HeraldVox）是一个 Coding Agent 的语音交互层。核心逻辑很简单：用户说话 → 语音识别 → 自然语言理解 → 指令下发给 Coding Agent → Agent 执行 → 结果通过 TTS 反馈给用户。

HeraldVox is a voice interface layer for Coding Agents. The core logic is simple: user speaks → speech recognition → natural language understanding → command to Coding Agent → Agent executes → result fed back via TTS.

问题：命令行把大部分人挡在了门外 The Problem: Terminals Lock Most People Out

Coding Agent 是目前使用 AI 最强大的方式——它能操作电脑、执行命令、写代码、自动化几乎一切。但它的默认界面是这样的：

Coding Agents are the most powerful way to use AI right now — they can operate your computer, run commands, write code, automate almost everything. But the default interface looks like this:

我家有两个小孩，一岁多和三岁多。大量时间花在陪孩子和做家务上，需要反复离开电脑。但每次给 Agent 下指令，我都必须坐下来，打开终端，打字。每次离开，Agent 就停在那里等我。

I have two kids, one just over 1 and one just over 3. I spend large chunks of my day with them, away from my desk. But every time I needed to give an Agent a new instruction, I had to sit down, open a terminal, type. Every time I left, the Agent just sat there waiting.

语音引擎：全链路自研 Voice Engine: Fully Self-Built Pipeline

这是整个产品技术复杂度最高的部分。从唤醒词到最终 TTS 输出，全链路都是自研的，不依赖任何第三方语音平台。

This is the most technically complex part of the product. From wake word to final TTS output, the entire pipeline is self-built — no third-party voice platform dependencies.

唤醒词检测采用本地离线方案，不依赖网络，激活前的语音完全不经过任何服务器。STT 针对技术术语做了专项优化——Coding Agent 场景下用户会说大量函数名、组件名、文件路径，通用 STT 识别率往往不够。TTS 采用流式方案，边生成边合成，显著降低首字延迟。

Wake word detection runs locally and offline — no audio reaches any server before activation. STT is optimized for technical terminology: in Coding Agent scenarios, users say many function names, component names, and file paths that standard STT handles poorly. TTS uses a streaming approach — synthesize as the text generates — which significantly reduces time-to-first-audio.

支持的 Agent Supported Agents

目前支持四个主流 Coding Agent，可以同时开多个并行运行，用语音一句话切换。

Currently supports four major Coding Agents. Run multiple in parallel, switch between them with a single voice command.

端到端加密通信 End-to-End Encrypted Communication

手机端控制桌面 Agent 的通信基于 libsodium，使用 X25519 密钥交换 + XSalsa20-Poly1305 对称加密。中继服务器只转发加密后的数据包，无法读取任何内容。

Mobile-to-desktop Agent communication uses libsodium: X25519 key exchange + XSalsa20-Poly1305 symmetric encryption. The relay server only forwards encrypted packets — it can't read anything.

怎么做到的：用 AI Agent 做 AI Agent How I Did It: Using AI Agents to Build AI Agents

这件事最有意思的地方是：我用来做唤达的工具，就是唤达要支持的那些工具。我的工作流是这样的：

The most interesting thing about this project: the tools I used to build HeraldVox are the exact tools it now supports. My workflow looked like this:

边带俩娃边创业。白天陪孩子，晚上写代码。有时候凌晨三点爬起来调 bug，六点又被小的那个哭醒。这个节奏持续了将近一年。整个产品的核心业务代码量超过了 20 万行。

Building while raising two kids. Days with them, nights writing code. Sometimes up at 3am debugging, back up at 6 when the younger one cried. That rhythm went on for nearly a year. The product's core business code now exceeds 200,000 lines.