Local QQ Voice Bot Build with NapCat and GPT-SoVITS
A first-person technical build record describes a local QQ voice chatbot using NapCat, AstrBot, an OpenAI-compatible model service and GPT-SoVITS. It covers a four-layer pipeline, reverse WebSocket setup, SILK v3 conversion through ffmpeg, zero-shot voice cloning with 3–10 seconds of clean reference audio, private and group memory, proactive-reply controls, latency tuning and layered troubleshooting. The report also says third-party protocol clients that use simulated login carry platform-rule, risk-control and possible account-ban risks. Its deployment experience and troubleshooting observations are not independently verified.
The record provides a concrete architecture and ordered validation method for a multi-component voice chatbot, showing where message delivery, model replies, speech synthesis, audio encoding, startup timing and speaking behavior can fail.
引子
文字聊天机器人已经很不稀奇了，但如果你想让自己的人设角色在群里用特定的声音说话，链路会突然长出一大截：协议端、机器人框架、大模型、语音合成，少一环都不行。
最近「大肥鱼」人设和「鲸鱼娘」桌宠在 DeepSeek 生态里火了一把，很多人第一次意识到：这些角色如果只会打字，乐趣少一半；真让它们用自己的声音在群里接话，才算「活过来」。
这篇文章是完整的一次搭建实录：从技术选型、端口规划，到语音编码绕不开的坑，再到把默认 TTS 换成支持零样本音色克隆的 GPT-SoVITS；最后还包括跑起来之后的那些问题——消息上下文怎么隔离、群消息历史留多少条、怎么让它主动接话又不招人烦、以及延迟该从哪一层下手。全程本地部署，不依赖任何商业语音服务。
一、先把架构拆成四层
语音聊天机器人不是「接个 LLM 就行了」，它是一条四段式的流水线：
#NapCat#AstrBot#GPT-SoVITS#DeepSeek#Edge-TTS#ffmpeg#OneBot 11