xeon_tts

Local TTS skill using OpenVINO Qwen3-TTS for voice cloning and emotion style synthesis, supporting QQBOT workflows with strict audio length and file retentio...

Security Scan

The skill's files and runtime instructions largely match a local TTS/voice-clone purpose, but the install/runtime scripts download code and models from external sources (including an unexpected HF mirror), install persistent systemd services, and run binaries — these choices are coherent with the feature set but introduce elevated risk and deserve manual review before install.

Purpose & Capability

Name/description (local OpenVINO Qwen3-TTS for QQBOT workflows) align with the included scripts and server: node workflow gateway, Flask TTS service, config injection into OpenClaw. The scripts only modify channels.qqbot.xeonTts and back up the OpenClaw config before writing, which matches the stated intent.

Instruction Scope

SKILL.md and scripts instruct the agent to install Python envs, system packages, register user-level systemd services, modify ~/.openclaw/openclaw.json, download models, and save uploaded audio to local references/ and outputs/. Those actions are within the described TTS workflow, but they include writing to user config, enabling autostart services, and making network calls — all of which are beyond a pure 'instruction-only' skill and should be permitted explicitly by the operator.

Install Mechanism

Installation downloads and executes software: Miniconda installer from repo.anaconda.com (expected), pip install of xdp-tts-service (PyPI by default), and model downloads via the Hugging Face CLI. Notably, setup_env.sh exports HF_ENDPOINT=https://hf-mirror.com and HF_HUB_ENABLE_HF_TRANSFER=0, which redirects HF operations to a third‑party mirror by default — that is unexpected and raises supply-chain risk. Model downloads and pip installs run code that will be written to disk and executed; review sources before proceeding.

Credentials

The skill does not declare or require any secret environment variables or cloud credentials. It does accept optional environment overrides (BASE_MODEL_REPO, CUSTOM_MODEL_REPO, XDP_TTS_PIP_SPEC, etc.) used to control model and package locations; these are appropriate for the task but are not enforced as required secrets.

Persistence & Privilege

The installer registers and enables two user-level systemd services (TTS Flask and Node gateway) and modifies the user's OpenClaw config (with a timestamped backup). Persistent autostart is justified for a local service but increases blast radius; the skill does not set always:true, but enabling services and writing config are persistent privileges the operator should approve.

What to consider before installing

This package looks functionally consistent with a local TTS/voice-clone skill, but you should not install it blindly. Before installing: 1) Inspect the xdp-tts-service package source (pip package) to verify it is trustworthy; 2) Check the model download sources — setup_env.sh sets HF_ENDPOINT to https://hf-mirror.com by default (an unexpected third‑party mirror). If you don't trust that mirror, override HF_ENDPOINT or ensure hf CLI points to official huggingface.co repos; 3) Review the full server.js (it spawns child processes and handles uploads) to see which external commands (ffprobe, etc.) it executes; 4) Be aware the installer creates user systemd services and will enable/start them and writes into ~/.openclaw/openclaw.json (a backup is created). If you want to reduce risk, run install steps in an isolated VM/container or run setup_env.sh with --skip-deps and inspect generated config files before enabling services. If anything looks unfamiliar (unknown HF mirror, unexpected pip package, or surprising exec calls), do not enable the services and seek a trusted upstream source.

基于 OpenVINO Qwen3-TTS Base/Custom 模型的本地语音合成技能，面向 OpenClaw 的 QQBOT 工作流使用。

目标

• 安装本地双服务：5002 Flask TTS，9002 Node TTS Workflow
• 自动配置目标机器自己的 OpenClaw 配置，但只写入 channels.qqbot.xeonTts
• 与 xeonasr 共存，不覆盖 tools.media.audio 或 channels.qqbot.stt
• 支持两个工作流：音色克隆、指定语气 TTS

什么时候应该调用 xeontts

只有在以下场景才使用 xeontts：

• 用户明确要“克隆音色”“克隆声音”“复制我的声音”
• 用户要求“用某种语气朗读/播报/生成语音”
• 用户要把音频生成到本地文件，而不是做转写

以下场景禁止走 xeontts：

• 识别语音
• 语音转文字
• 听写
• STT / ASR

这些请求必须交给 xeonasr，以避免任务冲突。

OpenClaw / QQBOT 使用规则

规则 1：音色克隆必须分两步走

当用户说“我要克隆音色”时：

1. 立即把当前会话切到 clone 流程
2. 回复用户上传 3 到 5 秒参考音频
3. 在收到参考音频前，不要直接开始合成
4. 如果机器上已安装 xeonasr，QQBOT 发来的语音会先命中 ASR；此时应由 ASR 把音频转交给 xeontts，而不是按普通转写处理
5. 收到音频后必须先校验时长
6. 如果时长小于 3 秒或大于 5 秒，直接拒绝并提示重新上传
7. 校验通过后，再让用户发送要朗读的文本
8. 使用 Base 模型生成音频并落盘

规则 2：指定语气生成默认走 Custom 模型

当用户说“用开心的语气朗读……”“生成语音……”时：

1. 解析用户是否指定了语气
2. 如果没有指定，默认使用 普通
3. 使用 Custom 模型生成音频
4. 把结果保存到本地 outputs/
5. 回复用户文件路径

规则 3：必须做时长保护

• 参考音频：3 到 5 秒
• Base 克隆输出：最多约 20 秒
• Custom 输出：最多约 30 秒

如果用户显式要求更长时长，或者文本长度估算会明显超过上限，必须直接提示用户拆分内容，而不是盲目提交推理。

规则 4：必须明确告知文件保留期

• 参考音频和生成结果默认只保留 7 天
• 超过保留期后，系统会自动清理 references/ 和 outputs/ 中的旧文件
• 在提示用户上传参考音频、以及告知生成完成时，都要明确告诉用户这一点

安装流程

当 OpenClaw、QQBOT 或其他代理被要求安装这个 skill 时，按下面流程执行：

1. 安装 skill

   clawhub install xeontts
   cd "$HOME/.openclaw/workspace/skills/xeontts"
   

2. 执行安装脚本

   bash install.sh
   

3. 安装脚本会完成这些事情

• 创建 Python 3.10 环境
• 从 PyPI 安装 xdp-tts-service
• 生成 config.json 与 tts_config.json
• 尝试下载 Base / Custom OV 模型
• 仅在显式提供旧兼容参数时，才额外下载 Base checkpoint
• 配置 OpenClaw 的 channels.qqbot.xeonTts
• 启动 5002 与 9002
• 注册用户级 systemd 服务
• 运行 self_check.sh

当前默认模型仓库：

• aurora2035/Qwen3-TTS-12Hz-0.6B-Base-OpenVINO-INT8
• aurora2035/Qwen3-TTS-12Hz-0.6B-CustomVoice-OpenVINO-INT8

运行时端口

服务	端口	作用
Flask TTS	5002	真正执行 TTS 推理
Node Workflow	9002	解析 QQBOT 任务、维护会话状态、校验音频/文本时长

OpenClaw 配置约定

xeontts 只会写入如下配置块：

{
  "channels": {
    "qqbot": {
      "xeonTts": {
        "enabled": true,
        "baseUrl": "http://127.0.0.1:9002",
        "cloneModel": "qwen3_tts_0.6b_base_openvino",
        "customModel": "qwen3_tts_0.6b_custom_openvino"
      }
    }
  }
}

这意味着：

• 不会覆盖现有 channels.qqbot.stt
• 不会动 tools.media.audio
• 不会和 xeonasr 抢同一条 STT 链路

常用命令

cd "$HOME/.openclaw/workspace/skills/xeontts"

bash start_all.sh
bash stop_tts.sh
bash self_check.sh
curl http://127.0.0.1:5002/api/health
curl http://127.0.0.1:9002/health

关键接口

• POST /api/workflow/message
  • 作用：根据用户消息判断是 clone 还是 custom TTS，或者提示补充参考音频
• POST /api/workflow/reference-audio
  • 作用：上传参考音频，校验 3 到 5 秒后入库
• POST /api/tts/custom-speak
  • 作用：直接调用 Custom 模型生成语音
• POST /api/tts/clone-speak
  • 作用：直接调用 Base 模型做音色克隆

故障排查

• 如果 5002 不通，先检查 tts.log
• 如果 9002 不通，先检查 skill.log
• 如果参考音频总是被拒绝，先确认机器上是否有可用的 ffprobe；当前版本对 WAV 参考音频也支持无 ffprobe 回退校验
• 如果用户说的是转写意图，不要误用 xeontts
• 如果 Base 模型报错，优先让用户更换更干净的 3 到 5 秒参考音频
• 当前默认发布形态只要求 Qwen3-TTS-12Hz-0.6B-Base-OpenVINO-INT8，不再默认要求原始 Base checkpoint
• 只有旧导出模型缺少 processor 或 speech tokenizer 权重时，才需要补 BASE_CHECKPOINT_PATH