Initial commit

BUSINESS-LOGIC.md

+# Luna TTS 排行榜冲榜 — 代码级业务逻辑梳理
+
+> 基于 `tts-ranking/` 代码实际实现，2026-07-09
+
+## 一、系统架构
+
+```
+Chrome 扩展 (MV3)
+├── content.js    → 页面注入、音频捕获、投票模拟、状态管理
+├── background.js → WebSocket 客户端、消息转发、Debugger API 原生输入
+├── inject.js     → fetch 拦截器（注入页面主世界）
+└── popup/        → 状态面板（启动/停止/统计）
+
+Python 服务
+├── server.py     → WebSocket 服务、音频下载、日志
+├── strategy.py   → 投票决策引擎
+├── matcher.py    → 声纹匹配引擎 (Resemblyzer)
+└── config.py     → 可配置参数
+```
+
+## 二、对照音频库
+
+三个库，通过 subdirectory symlinks 组织：
+
+| 库 | 目录 | 厂商数 | 用途 |
+|---|------|--------|------|
+| 我方 | `standard/` | 1 (VUI Labs) | 识别自己的音频，投自己 |
+| 大厂 | `big_vendors/` | 4 (SpeechifyAI/Google/Cartesia/Alibaba) | 优先投票目标 |
+| 竞品 | `competitors/` | 16 | 仅监控+兜底决策 |
+
+每个 subdirectory 命名格式：`Vendor__Model`，内部包含该模型的多个 mp3 文件。
+
+## 三、声纹匹配流程
+
+1. **启动时加载**：遍历三个目录，对每个 mp3 文件：
+   - `preprocess_wav()` → 音频预处理（静音裁剪）
+   - `VoiceEncoder.embed_utterance()` → 256 维声纹向量
+   - 文件名提取 speaker 名（兼容新旧两种格式）
+   - 文件名/pitch 判断性别
+   - 同 speaker 多片段取最高分
+
+2. **运行时匹配**（收到音频 URL）：
+   - 服务端下载音频 → 提取声纹
+   - pitch 检测未知音频性别 → 只与同性别对照音频匹配
+   - 每个库返回：best speaker + best score + 全量分数 + vendor/model
+   - 三个库各自的最佳结果汇总为 SideResult
+
+3. **设备自动检测**：CUDA → MPS → CPU
+
+## 四、投票决策引擎 (`strategy.py`)
+
+### 核心逻辑
+
+对左右两侧音频，分别从三个库的全量匹配中取**跨库最高分**，确定该侧的 `top_library` 和 `top_speaker`。然后按优先级决策：
+
+| 优先级 | 条件 | 决策 |
+|--------|------|------|
+| 1 | 某侧 top_library == "our" AND top_score >= 0.80 | 投该侧 |
+| 1a | 两侧都是 ours | 投高分侧（平局投左） |
+| 2 | 某侧 top_library == "big_vendor" AND top_score >= 0.80 | 投该侧 |
+| 2a | 两侧都是 big_vendor | 随机投一侧 |
+| 3a | 两侧 top_library == "competitor" AND >= 0.80 | **刷新页面**（弃权） |
+| 3b | 一侧 competitor，另一侧不是 | **投非竞品那一侧** |
+| 3c | 都不是 competitor | **随机投一侧** |
+
+### 可配置参数 (`config.py`)
+
+| 参数 | 默认值 | 说明 |
+|------|--------|------|
+| `OUR_THRESHOLD` / 匹配阈值 | 0.80 | 所有库共用一个阈值 |
+| `DAILY_VOTE_LIMIT` | 800 | 达到后停止 |
+| `WS_PORT` | 8765 | WebSocket 端口 |
+
+（注：`COMPETITOR_THRESHOLD` 和 `BIG_VENDOR_VOTE_PROBABILITY` 仍存在于 config.py 但 server.py 已不再引用）
+
+## 五、扩展端工作流 (`content.js`)
+
+### 5.1 初始化
+- `document_start` 注入 `inject.js` 劫持 `window.fetch`，捕获 mp3/wav 请求 URL
+- 自动恢复：`chrome.storage.local` 存 enabled 状态，页面刷新后 3 秒自动续投
+
+### 5.2 单轮投票循环 (`doVote`)
+
+```
+1. 检测 CAPTCHA（最多 20s）
+   - 自动点击 Turnstile（Debugger API 原生点击）
+   - 超时 → 刷新页面
+2. 捕获 2 个音频 URL（最多 60s）
+   - 优先读 fetch 拦截队列
+   - 兜底读 <audio> DOM
+   - 使用消费指针 _consumedUrls 区分新旧音频
+3. 等投票卡片渲染（最多 30s）
+4. grant_autoplay（Debugger API 原生点击获取用户激活）
+5. 点击播放按钮 + audio.play()
+6. 等音频播放（最少 3s，最多 8s）
+7. 发送 match_request → 等待匹配结果
+8. 根据决策执行投票：
+   - vote_left/right → simulateVote → 验证 → 标记完成
+   - skip → skipAndRefresh（等 20s 后刷新）
+9. scheduleNextVote（2-8s 随机延迟）
+```
+
+### 5.3 投票模拟 (`simulateVote`)
+- 主方式：Debugger API 发送原生 keydown/keyup（ArrowLeft/ArrowRight）
+- 失败兜底：synthetic keyboard + card click（实际无效）
+- 验证：等 5s，检查是否有新 URL 入队；无则重试一次等 5s；再失败标为 FAILED → doVote 刷新页面
+
+### 5.4 Reveal 验证 (`_captureRevealAndVerify`)
+- 投票前启动 200ms 轮询
+- 页面揭示后抓到 `.animate-arena-upvote`（投票的模型名）和 `.animate-arena-downvote`（另一个模型名）
+- 直接输出到浏览器控制台（带颜色标注 + 对比结果 ✓/✗）
+
+### 5.5 工作/休息周期
+- 工作 10 分钟 → 休息 60-180 秒（随机）→ 继续工作
+
+### 5.6 反检测
+- 聆听后投票（min 3s）
+- 投票间隔 2-8s 随机
+- 弃权刷新：前 9 次随机 1-3s，第 10 次冷却 10-20s（通过 storage 跨页面持久化计数）
+- Debugger API 原生输入事件（isTrusted: true）
+
+## 六、通信协议
+
+WebSocket JSON（localhost:8765）
+
+### match_request（扩展 → 服务端）
+```json
+{
+  "type": "match_request",
+  "request_id": "uuid",
+  "audio_left": "https://cdn.../audio_a.mp3",
+  "audio_right": "https://cdn.../audio_b.mp3"
+}
+```
+
+### match_result（服务端 → 扩展）
+```json
+{
+  "type": "match_result",
+  "decision": "vote_left|vote_right|skip",
+  "reason": "决策原因",
+  "details": {
+    "left": { "our_match": ..., "competitor_match": ..., "big_vendor_match": ..., "top_match": {...} },
+    "right": { ... }
+  },
+  "stats": { "today_votes": 23, "today_skips": 7, "daily_limit": 800 }
+}
+```
+
+### 控制消息
+- `status` — 查询当日统计
+- `grant_autoplay` — 请求原生点击
+- `dispatch_key` — 请求原生键盘事件
+- `click_at` — 请求原生鼠标点击
+- `captcha_detected` — 通知 CAPTCHA
+
+## 七、日志格式
+
+每次匹配输出到终端 + `server/logs/match-YYYYMMDD.log`：
+
+```
+───────────────────────────────────────────────────────
+  LEFT 音频
+  我方最佳: Ella(0.7863)
+    Ella: 0.7863
+    Grace: 0.6211
+
+  竞品最佳: Lively Girl_us_female(0.8549)
+    Lively Girl_us_female: 0.8549 ←  [ElevenLabs / Eleven_v3]
+    ...
+
+  大厂最佳: Caroline_us_female(0.8592)
+    Caroline_us_female: 0.8592 ←  [Cartesia / Sonic_3.5]
+
+  >>> left 最高: Caroline_us_female(0.8592) [big_vendor]
+
+  RIGHT 音频
+  ...
+
+  决策: vote_left
+  原因: left 最高分为大厂 Caroline_us_female(0.8592) → vote
+  统计: 23 票 / 7 弃
+───────────────────────────────────────────────────────
+```
+
+标注说明：`←` 表示达到阈值 0.80；`[厂商 / 模型]` 表示所属厂商和模型。
+
+## 八、竞品模型清单（当前加载的）
+
+**大厂 (4)**：SpeechifyAI/Simba_3.2、Google/Gemini_3.1_Flash_TTS、Cartesia/Sonic_3.5、Alibaba/Fun-Realtime-TTS
+
+**竞品 (19)**：Inworld（3个模型）、SpaceXAI、MiniMax（4个模型）、async（2个模型）、StepFun（3个模型）、ElevenLabs、Smallest.ai、Fish Audio（2个模型）、Microsoft、SpeechifyAI/Simba_3.0（竞品版本）

README.md

+# Luna TTS 排行榜冲榜工具
+
+Native Messaging 方案，扩展直连本地 Python 进程，不走网络层。兼容代理环境（AdsPower / Quark / SunBrowser 等指纹浏览器）。
+
+## Mac 安装
+
+### 0. 前置条件
+
+```bash
+# 1. 确认 Python 3.8+
+python3 --version
+
+# 2. 安装 Xcode Command Line Tools（提供 gcc 编译器）
+xcode-select --install
+
+# 3. 安装 Python 依赖
+cd 项目路径/tts-ranking
+python3 -m pip install -r requirements.txt
+```
+
+### 1. 加载扩展
+
+打开 Chrome（或指纹浏览器），地址栏输入 `chrome://extensions/` → 开「开发者模式」→ 「加载已解压的扩展程序」→ 选择 `extension/` 目录。
+
+记住卡片上显示的 **扩展 ID**（一长串字母）。
+
+### 2. 安装 Native Host
+
+```bash
+cd 项目路径/tts-ranking/server
+./install_host.sh 刚才的扩展ID
+```
+
+脚本会自动编译 wrapper、安装到 Chrome/Chromium/Quark/Edge/AdsPower 所有浏览器目录。
+
+> 如果报 `gcc: command not found`，先跑 `xcode-select --install`。
+
+> 每次移除扩展再重新加载，ID 会变，需要重新跑这步。
+
+### 3. 刷新扩展
+
+`chrome://extensions/` → 点 Luna TTS Bot 的刷新 ⟳。
+
+### 4. 验证
+
+打开 `https://artificialanalysis.ai/text-to-speech/arena`，点扩展图标，应显示绿色 `Connected`。
+
+---
+
+## Windows 安装
+
+### 1. 安装 Python 依赖
+
+```powershell
+cd C:\path\to\tts-ranking
+python -m pip install -r requirements.txt
+```
+
+### 2. 编译 Native Host
+
+需要 gcc（可用 MinGW）或 Visual Studio。
+
+```powershell
+cd server
+gcc -o run_host.exe runner_win.c
+```
+
+### 3. 加载扩展
+
+`chrome://extensions/` → 开发者模式 → 加载已解压 → 选 `extension/` 目录。
+
+### 4. 注册 Native Host
+
+创建 `install_host.reg`（用记事本，注意把路径改成实际项目路径）：
+
+```reg
+Windows Registry Editor Version 5.00
+
+[HKEY_CURRENT_USER\Software\Google\Chrome\NativeMessagingHosts\com.luna.tts]
+@="C:\\path\\to\\tts-ranking\\server\\run_host.exe"
+```
+
+双击导入注册表。
+
+指纹浏览器（AdsPower/Quark）需要同时导入到对应注册表路径。
+
+### 5. 刷新扩展并验证
+
+刷新扩展 → 打开打榜页面 → 检查弹窗是否绿了。
+
+---
+
+## 使用
+
+1. 打开 `https://artificialanalysis.ai/text-to-speech/arena`
+2. 有人机校验就手动点一下
+3. 点扩展图标 → **Start Voting**
+4. 页面右下角绿色提示条显示当前状态
+
+## 人工介入
+
+两侧都匹配到我方，同时有竞品也匹配到时，系统暂停等待。按快捷键：
+
+| 快捷键 | 操作 |
+|--------|------|
+| `Ctrl+Shift+L` | 强制投左 |
+| `Ctrl+Shift+R` | 强制投右 |
+| `Ctrl+Shift+S` | 跳过本轮 |
+
+5 分钟无操作自动刷新页面。
+
+## 日志
+
+- 匹配日志：`server/logs/native-YYYYMMDD.log`
+- 崩溃日志：`server/logs/crash.log`
+- 扩展状态：`chrome://extensions/` → Luna TTS Bot → 点 service worker 链接 → Console
+
+## 故障排查
+
+### 扩展显示 Disconnected
+
+1. 检查 `server/logs/crash.log`，看是否有 `ModuleNotFoundError`（依赖没装全）
+2. 确认 `install_host.sh` 成功跑过，`com.luna.tts.json` 在各浏览器目录下都存在：
+   ```bash
+   find ~/Library -name "com.luna.tts.json" 2>/dev/null
+   ```
+3. 确认 manifest 里的扩展 ID 和 `chrome://extensions/` 上的一致：
+   ```bash
+   cat ~/Library/Application\ Support/Google/Chrome/NativeMessagingHosts/com.luna.tts.json
+   ```
+4. 完全退出浏览器再重新打开，刷新扩展
+
+### 稳定连接但报 `debugger blocked`
+
+指纹浏览器可能限制了 Debugger API。CAPTCHA 自动点击和键盘投票会降级为备用方式，如果不是经常出现可以忽略。
+
+### 连上又断开、反复循环
+
+看 `server/logs/crash.log`——大概率是依赖问题或 Python 版本问题。确认：
+
+```bash
+python3 -c "import torch; import torchaudio; import resemblyzer; print('OK')"
+```
+
+### 跨芯片兼容（Apple Silicon ↔ Intel）
+
+编译产物 `run_host` 绑定当前机器的芯片架构。发给别人用前，确认对方机器架构：
+
+```bash
+# 对方机器跑这个看芯片
+uname -m
+# arm64  = Apple Silicon (M1/M2/M3/M4)
+# x86_64 = Intel Mac
+```
+
+如果不匹配（比如你在 M 芯片编译、发给 Intel 机器），对方会报 `bad CPU type in executable`。解决：让对方面重新编译：
+
+```bash
+cd 项目路径/tts-ranking/server
+gcc -o run_host runner.c
+```
+
+### 提示需要手动点击人机验证
+
+Cloudflare Turnstile PAT 挑战无法自动绕过，页面右下角会提示"人机验证，需要手动点击"，点一下即可。

download_tts.py

+#!/usr/bin/env python3
+"""Download remaining TTS models from Artificial Analysis (batch 2)."""
+import subprocess
+import json
+import os
+import time
+import re
+import sys
+
+sys.stdout.reconfigure(line_buffering=True) if hasattr(sys.stdout, 'reconfigure') else None
+
+BASE_URL = "https://artificialanalysis.ai/api/text-to-speech/speech-explorer"
+OUTPUT_DIR = "/Users/chengyubing/Documents/projects/hzjj/tts-ranking/server/reference/raw_audios"
+
+MODELS = [
+    ("Speech-02-Turbo", "953aeced-7f15-4a02-a2f5-3c9cfccbbb39", "MiniMax"),
+    ("Magpie Multilingual", "04719149-11ed-42b1-bc2c-166860c03c23", "Magpie"),
+    ("Magpie-Multilingual 357M (Feb 2026)", "49198771-e5ea-4c4e-bad9-e57671a92291", "Magpie"),
+    ("Coda", "cddf7ab5-43d6-488f-be03-1dcfbb8955d3", "Coda"),
+    ("SSFM-v30", "b86c54d8-b51d-4186-b412-a705f2b9b76d", "SSFM"),
+    ("Qwen3 TTS", "8eb42b0b-213a-4911-b096-a7455e944090", "Alibaba"),
+    ("Seed-TTS 2.0", "31ef1e00-542b-4d06-8b0f-0bb52ef4e518", "ByteDance"),
+    ("GPT-Realtime-2", "99365f44-415e-4ff2-94c9-59b8b2ec5847", "OpenAI"),
+    ("Polly Standard", "d3f9b298-8e1a-4328-a2bd-ce3ceea6db1f", "Amazon"),
+    ("SSFM-v21", "c8aa7fcf-69c0-4c7f-9ead-f640801a8069", "SSFM"),
+    ("Voxtral TTS", "b424cd2e-abb9-4131-8f47-7731bdbff455", "Voxtral"),
+    ("MiMo-V2.5-TTS", "abdcf494-6c73-4574-ade9-6672cdb21936", "MiMo"),
+    ("Octave 2", "b4a9f921-ae13-4c12-b907-6e9c6bedaf58", "Octave"),
+    ("Gemini 2.5 Flash TTS (Dec 2025)", "7b6ea07a-99e1-4635-9ea1-beee1b6520d5", "Google"),
+    ("Arcana v3", "20c6b2a3-4881-4f5d-b82f-f135fab82711", "Arcana"),
+    ("Higgs Audio V3 TTS", "49dc49c3-828e-4af5-8527-49606b4f6e14", "Higgs Audio"),
+    ("Zonos-v0.1", "f5a3ab85-032b-41b1-8c78-c4b0eef2a9e8", "Zyphra"),
+    ("Fun-Realtime-TTS-Preview", "48100868-7939-4cb3-a5be-8fb1202061ec", "Alibaba"),
+    ("Standard", "7efc9790-17f6-44a5-8b4f-897deaeb0088", "MiniMax"),
+    ("MetaVoice v1", "fb51ab93-d79a-49c5-a2f5-19f03e311afd", "Meta"),
+    ("Gradium TTS", "2f3dbd41-a9c4-4f0c-a16b-427dea9feb6e", "Gradium"),
+    ("MiMo-V2-TTS", "4489b587-5602-434a-9c24-772df19e879b", "MiMo"),
+    ("MAI-Voice-1", "c8b0b07c-26a7-444c-a8ff-c412a3d433b7", "MAI"),
+    ("XTTS v2", "f2f2989e-69d1-4b1c-a419-ee316e6c1c3d", "Coqui"),
+    ("Speech 2.6", "6d2c7daa-7585-4b85-b1db-239c1c29b9a3", "MiniMax"),
+    ("WaveNet", "16ad3e89-cc78-4478-a5a0-a048b5c72cfb", "Google"),
+    ("Speech-02-HD", "65166a7b-b42c-4a68-b051-9662facd2f91", "MiniMax"),
+    ("Qwen3 TTS Flash", "f05bc243-0723-4f05-aa19-31341d8f2677", "Alibaba"),
+    ("Polly Neural", "eb3b0d75-134b-4e6a-82c5-4db2398d5c07", "Amazon"),
+    ("Falcon (Beta)", "5ab04a8d-c698-4726-bebd-af5fc553b752", "Falcon"),
+    ("Octave TTS", "009f6201-1862-454e-87b5-e2a367bc027f", "Octave"),
+    ("Neural2", "666ff702-a2c8-413a-a8df-61902a6e7cb9", "Google"),
+    ("Neuphonic TTS", "55912533-f2de-48e4-8318-990c435f1401", "Neuphonic"),
+    ("VibeVoice 7B", "7395ed9e-1d19-4ab8-a989-eca165beefa4", "VibeVoice"),
+    ("Orion", "90c81fc8-01ce-4e1d-bcd4-a57dff1c25cf", "Orion"),
+    ("Chatterbox", "6cb43d71-bd49-4b44-9635-9ea5f28fd376", "Chatterbox"),
+    ("VibeVoice 1.5B", "bd053b0f-5f69-4313-a7e5-5acd71422d4a", "VibeVoice"),
+    ("SIMBA 1.6", "e0ca6a5b-6888-4e0f-8f4e-2c4cb59dd2aa", "SpeechifyAI"),
+    ("Gemini 2.5 Flash Lite TTS", "a3824338-576c-4deb-8446-264eb9895db2", "Google"),
+    ("Multilingual v2", "9cd43eb4-f436-4ea2-b3d9-5c912bdea912", "ElevenLabs"),
+    ("T2A-01-HD", "0707d10e-e111-484c-8fd5-5dbb0a78029e", "MiniMax"),
+    ("OpenAudio S1", "46730007-0d7b-4395-8077-2b5104ea79b0", "OpenAudio"),
+    ("Turbo v2.5", "2d3f6b08-353d-40e9-a5e7-a3c1952b2dd8", "ElevenLabs"),
+    ("Sonic 3", "a71d7774-e73e-4b6c-8dfe-7963273b2975", "Cartesia"),
+    ("Maya1", "081709c5-d010-462f-b6ad-ef128df78cbd", "Maya"),
+    ("T2A-01-Turbo", "c8c8fcf1-03cb-4b63-a6c9-d72cab227b3f", "MiniMax"),
+    ("Fish Speech 1.5", "8bdfce7b-e8c5-4a29-984a-bd360d913c31", "Fish Audio"),
+    ("Chatterbox HD", "06b8f3e3-e4fe-4cf7-9774-30816972bb59", "Chatterbox"),
+    ("Polly Long-Form", "8b4be309-2f31-4ff2-be79-a6652c29ab12", "Amazon"),
+    ("Murf Speech Gen 2", "86d3b629-8222-4ab6-ba77-6d2841fbf212", "Murf"),
+    ("Mist V2", "5d922a1e-af77-4dae-adb4-7428dd634d5a", "Mist"),
+    ("StyleTTS 2", "3b322e37-a016-4b2e-aacb-1aef8df00829", "StyleTTS"),
+    ("Kokoro 82M v1.0", "6de03a42-66ce-484a-9c01-95fe9b4d7422", "Kokoro"),
+    ("Flash v2.5", "cccb3742-da05-4447-910c-43a5d3eda5b7", "ElevenLabs"),
+    ("Studio", "02265a20-67bc-4892-89d0-f54545b6b4a1", "ElevenLabs"),
+    ("Async Flash v1.0", "765fd2b5-acb4-490f-a402-eaad710f23e1", "async"),
+    ("TTS-1", "78588185-b188-46fb-8d06-c06f52009715", "OpenAI"),
+    ("SIMBA 1.0", "588357bb-fe98-4286-b13e-3cae200e6085", "SpeechifyAI"),
+    ("TTS-1 HD", "4a2776b8-a0c6-42f2-8ebe-db364f1ddf13", "OpenAI"),
+    ("Polly Generative", "4ef13968-15ee-4311-a6d1-685f015b0717", "Amazon"),
+    ("Journey", "2b5bb443-e32e-4cb4-a9f2-918c00210211", "Journey"),
+    ("Chirp 3: HD", "f524429e-7e12-47eb-88df-6a9730f88989", "Google"),
+    ("Gemini 2.5 Pro (Dec 2025)", "2e9ecf4e-7aca-41ac-9436-a5c0c5664fe0", "Google"),
+    ("LMNT", "a8b4b1db-2a38-4649-9b57-a12fb0a85253", "LMNT"),
+    ("OpenVoice v2", "80791fbe-7454-4719-9063-bedfcde606c7", "OpenVoice"),
+    ("Fish Audio S1", "cfb0a07b-94b8-4b05-8d2c-3529f66cd1d2", "Fish Audio"),
+    ("Sonic English (Oct 2024)", "798934ab-0f43-49d2-8343-15c1ade78458", "Cartesia"),
+]
+
+
+def safe_filename(name):
+    return re.sub(r'[<>:"/\\|?*()]', '', name).replace(' ', '_').strip()
+
+
+def fetch_page(model_id, offset):
+    url = f"{BASE_URL}?model={model_id}&offset={offset}"
+    result = subprocess.run(
+        ['curl', '-s', '--max-time', '30', url],
+        capture_output=True, text=True, timeout=35
+    )
+    if result.returncode != 0 or not result.stdout:
+        raise RuntimeError(f"API call failed")
+    return json.loads(result.stdout)
+
+
+def download_file(url, dest):
+    if os.path.exists(dest) and os.path.getsize(dest) > 100:
+        return True
+    result = subprocess.run(
+        ['curl', '-s', '-L', '--max-time', '60', '-o', dest, url],
+        capture_output=True, text=True, timeout=65
+    )
+    if result.returncode != 0:
+        return False
+    if os.path.getsize(dest) < 100:
+        os.remove(dest)
+        return False
+    return True
+
+
+def download_model(model_name, model_id, creator):
+    folder_name = safe_filename(model_name)
+    folder_path = os.path.join(OUTPUT_DIR, creator, folder_name)
+    os.makedirs(folder_path, exist_ok=True)
+
+    existing = len([f for f in os.listdir(folder_path) if f.endswith('.mp3')])
+
+    try:
+        data = fetch_page(model_id, 0)
+    except Exception as e:
+        print(f"  [{creator}] {model_name}: ERROR {e}", flush=True)
+        return 0, 0
+
+    total = data['total']
+    if total == 0:
+        print(f"  [{creator}] {model_name}: 0 files, skip", flush=True)
+        return 0, 0
+
+    if existing >= total:
+        print(f"  [{creator}] {model_name}: {existing}/{total} (done)", flush=True)
+        return existing, 0
+
+    print(f"  [{creator}] {model_name}: {existing}/{total}, downloading...", flush=True)
+
+    downloaded = existing
+    failed = 0
+    offset = 0
+
+    while True:
+        try:
+            if offset > 0:
+                data = fetch_page(model_id, offset)
+                time.sleep(0.3)
+
+            for audio in data['audios']:
+                file_url = audio['fileName']
+                voice_name = audio['voice']['name']
+                accent = audio['voice']['accent']['slug']
+                gender = audio['voice']['gender']
+                file_id = file_url.split('/')[-1]
+                dest = os.path.join(folder_path, f"{voice_name}_{accent}_{gender}_{file_id}")
+                if download_file(file_url, dest):
+                    downloaded += 1
+                else:
+                    failed += 1
+
+            has_more = data['has_more']
+            offset = data['offset']
+            if not has_more:
+                break
+        except Exception as e:
+            print(f"    ERROR at offset {offset}: {e}", flush=True)
+            offset += 20
+            if offset >= total:
+                break
+            time.sleep(2)
+
+    print(f"    Done: {downloaded} ok, {failed} fail (total {total})", flush=True)
+    return downloaded, failed
+
+
+def main():
+    os.makedirs(OUTPUT_DIR, exist_ok=True)
+    print(f"Downloading {len(MODELS)} remaining models to {OUTPUT_DIR}", flush=True)
+
+    total_ok = 0
+    total_fail = 0
+    for i, (model_name, model_id, creator) in enumerate(MODELS, 1):
+        ok, fail = download_model(model_name, model_id, creator)
+        total_ok += ok
+        total_fail += fail
+        time.sleep(0.5)
+
+    print(f"\nALL DONE: {total_ok} ok, {total_fail} failed", flush=True)
+
+
+if __name__ == '__main__':
+    main()

extension/background.js

+// Luna TTS Ranking Bot — Background Service Worker
+// Bridges content script ↔ Python native host via Chrome Native Messaging
+
+const NATIVE_HOST = 'com.luna.tts';
+
+let port = null;
+let requestCallbacks = new Map();
+let latestStats = { today_votes: 0, today_skips: 0, daily_limit: 800 };
+let isConnected = false;
+
+// ---- Native Messaging lifecycle ----
+
+function connect() {
+  if (port) { try { port.disconnect(); } catch (e) {} }
+
+  port = chrome.runtime.connectNative(NATIVE_HOST);
+
+  port.onMessage.addListener((msg) => {
+    if (msg.type === 'match_result' && msg.request_id) {
+      const callback = requestCallbacks.get(msg.request_id);
+      if (callback) { callback(msg); requestCallbacks.delete(msg.request_id); }
+    }
+    if (msg.stats) latestStats = msg.stats;
+    chrome.tabs.query({ active: true, currentWindow: true }, (tabs) => {
+      if (tabs[0]) chrome.tabs.sendMessage(tabs[0].id, { type: 'from_server', payload: msg }).catch(() => {});
+    });
+    chrome.runtime.sendMessage({ type: 'from_server', payload: msg }).catch(() => {});
+  });
+
+  port.onDisconnect.addListener(() => {
+    console.log('[bg] Native host disconnected');
+    isConnected = false;
+    broadcastState('disconnected');
+    port = null;
+    setTimeout(connect, 5000);
+  });
+
+  isConnected = true;
+  console.log('[bg] Connected to native host');
+  broadcastState('connected');
+}
+
+connect();
+
+// ---- Message routing ----
+
+chrome.runtime.onMessage.addListener((message, sender, sendResponse) => {
+  if (message.type === 'match_request') {
+    if (port && isConnected) {
+      requestCallbacks.set(message.request_id, sendResponse);
+      port.postMessage({ type: 'match_request', request_id: message.request_id, audio_left: message.audio_left, audio_right: message.audio_right });
+    } else {
+      sendResponse({ type: 'match_result', decision: 'skip', reason: 'Python host not connected', details: {} });
+    }
+    return true;
+  }
+  if (message.type === 'get_state') {
+    sendResponse({ ws_connected: isConnected, ...latestStats });
+    return true;
+  }
+  if (message.type === 'verify_result' && port && isConnected) {
+    port.postMessage(message);
+    return false;
+  }
+});
+
+function broadcastState(status) {
+  chrome.tabs.query({ active: true, currentWindow: true }, (tabs) => {
+    if (tabs[0]) chrome.tabs.sendMessage(tabs[0].id, { type: 'ws_state', status }).catch(() => {});
+  });
+}
+
+// ---- Native input dispatch via Chrome Debugger API ----
+
+async function _attachDebugger(tabId) {
+  return new Promise(resolve => {
+    chrome.debugger.attach({ tabId }, '1.3', () => {
+      if (chrome.runtime.lastError) { console.log('[bg] Debugger attach failed:', chrome.runtime.lastError.message); resolve(false); }
+      else resolve(true);
+    });
+  });
+}
+
+async function _debuggerSend(tabId, method, params) {
+  return new Promise(resolve => {
+    chrome.debugger.sendCommand({ tabId }, method, params, (result) => {
+      resolve(chrome.runtime.lastError ? null : result);
+    });
+  });
+}
+
+chrome.runtime.onMessage.addListener((message, sender, sendResponse) => {
+  const tabId = sender.tab?.id;
+
+  if (message.type === 'grant_autoplay') {
+    (async () => {
+      if (!tabId) { sendResponse({ ok: false }); return; }
+      if (!await _attachDebugger(tabId)) { sendResponse({ ok: false, error: 'debugger blocked' }); return; }
+      await _debuggerSend(tabId, 'Input.dispatchMouseEvent', { type: 'mousePressed', x: message.x || 500, y: message.y || 400, button: 'left', clickCount: 1 });
+      await _debuggerSend(tabId, 'Input.dispatchMouseEvent', { type: 'mouseReleased', x: message.x || 500, y: message.y || 400, button: 'left', clickCount: 1 });
+      chrome.debugger.detach({ tabId }, () => {});
+      sendResponse({ ok: true });
+    })();
+    return true;
+  }
+
+  if (message.type === 'click_at') {
+    (async () => {
+      if (!tabId) { sendResponse({ ok: false }); return; }
+      if (!await _attachDebugger(tabId)) { sendResponse({ ok: false }); return; }
+      await _debuggerSend(tabId, 'Input.dispatchMouseEvent', { type: 'mouseMoved', x: message.x, y: message.y });
+      await new Promise(r => setTimeout(r, 100 + Math.random() * 200));
+      await _debuggerSend(tabId, 'Input.dispatchMouseEvent', { type: 'mousePressed', x: message.x, y: message.y, button: 'left', clickCount: 1 });
+      await new Promise(r => setTimeout(r, 50 + Math.random() * 100));
+      await _debuggerSend(tabId, 'Input.dispatchMouseEvent', { type: 'mouseReleased', x: message.x, y: message.y, button: 'left', clickCount: 1 });
+      chrome.debugger.detach({ tabId }, () => {});
+      sendResponse({ ok: true });
+    })();
+    return true;
+  }
+
+  if (message.type === 'dispatch_key') {
+    (async () => {
+      if (!tabId) { sendResponse({ ok: false }); return; }
+      if (!await _attachDebugger(tabId)) { sendResponse({ ok: false }); return; }
+      const key = message.key === 'left' ? 'ArrowLeft' : 'ArrowRight';
+      const vk = message.key === 'left' ? 37 : 39;
+      await _debuggerSend(tabId, 'Input.dispatchKeyEvent', { type: 'rawKeyDown', key, code: key, windowsVirtualKeyCode: vk });
+      await _debuggerSend(tabId, 'Input.dispatchKeyEvent', { type: 'keyUp', key, code: key, windowsVirtualKeyCode: vk });
+      chrome.debugger.detach({ tabId }, () => {});
+      sendResponse({ ok: true });
+    })();
+    return true;
+  }
+});
+
+chrome.alarms.create('keepalive', { periodInMinutes: 1 });
+chrome.alarms.onAlarm.addListener(() => {});

extension/inject.js

+// Injected into page main world — intercepts fetch for audio URLs
+(() => {
+  const _fetch = window.fetch;
+  window.fetch = function (...args) {
+    const url = typeof args[0] === 'string' ? args[0] : (args[0] && args[0].url) || '';
+    if (/\.(mp3|wav|ogg|m4a|flac)(\?|#|$)/i.test(url)) {
+      window.postMessage({ type: '__luna_audio', url }, '*');
+    }
+    return _fetch.apply(this, args);
+  };
+})();

extension/manifest.json

+{
+  "manifest_version": 3,
+  "name": "Luna TTS Ranking Bot",
+  "version": "0.1.0",
+  "description": "Semi-automated voting on Luna TTS Arena with voiceprint matching",
+  "permissions": [
+    "storage",
+    "alarms",
+    "activeTab",
+    "scripting",
+    "debugger",
+    "nativeMessaging"
+  ],
+  "host_permissions": [
+    "https://artificialanalysis.ai/*"
+  ],
+  "background": {
+    "service_worker": "background.js"
+  },
+  "content_scripts": [
+    {
+      "matches": ["https://artificialanalysis.ai/text-to-speech/arena*"],
+      "js": ["content.js"],
+      "run_at": "document_start"
+    }
+  ],
+  "action": {
+    "default_popup": "popup/popup.html",
+    "default_title": "Luna TTS Bot"
+  },
+  "icons": {},
+  "web_accessible_resources": [
+    {
+      "resources": ["inject.js"],
+      "matches": ["https://artificialanalysis.ai/*"]
+    }
+  ]
+}

extension/popup/popup.html

+<!DOCTYPE html>
+<html>
+<head>
+  <meta charset="UTF-8">
+  <style>
+    * { box-sizing: border-box; margin: 0; padding: 0; }
+    body { width: 280px; font-family: -apple-system, BlinkMacSystemFont, sans-serif; padding: 16px; font-size: 13px; }
+    h2 { font-size: 15px; margin-bottom: 12px; }
+    .status { display: flex; align-items: center; gap: 6px; margin-bottom: 12px; }
+    .dot { width: 8px; height: 8px; border-radius: 50%; display: inline-block; }
+    .dot.on { background: #22c55e; }
+    .dot.off { background: #ef4444; }
+    .row { display: flex; justify-content: space-between; margin-bottom: 6px; }
+    .label { color: #6b7280; }
+    .value { font-weight: 600; }
+    button { width: 100%; padding: 8px; border: none; border-radius: 6px; font-size: 13px; cursor: pointer; margin-top: 12px; }
+    button.start { background: #22c55e; color: white; }
+    button.stop  { background: #ef4444; color: white; }
+    button:disabled { opacity: 0.5; cursor: default; }
+    hr { border: none; border-top: 1px solid #e5e7eb; margin: 12px 0; }
+    .log { font-size: 11px; color: #6b7280; max-height: 100px; overflow-y: auto; }
+  </style>
+</head>
+<body>
+  <h2>Luna TTS Bot</h2>
+  <div class="status">
+    <span class="dot" id="dot"></span>
+    <span id="statusText">Disconnected</span>
+  </div>
+
+  <div class="row"><span class="label">Server</span><span class="value" id="serverStatus">—</span></div>
+  <div class="row"><span class="label">Today votes</span><span class="value" id="todayVotes">0 / 800</span></div>
+  <div class="row"><span class="label">Today skips</span><span class="value" id="todaySkips">0</span></div>
+
+  <button id="btnStart" class="start">Start Voting</button>
+  <button id="btnStop"  class="stop"  disabled>Stop</button>
+
+  <hr>
+  <div class="log" id="log"></div>
+
+  <script src="popup.js"></script>
+</body>
+</html>

extension/popup/popup.js

+// Luna TTS Ranking Bot — Popup controller
+
+const btnStart = document.getElementById('btnStart');
+const btnStop = document.getElementById('btnStop');
+const statusText = document.getElementById('statusText');
+const dot = document.getElementById('dot');
+const serverStatus = document.getElementById('serverStatus');
+const todayVotes = document.getElementById('todayVotes');
+const todaySkips = document.getElementById('todaySkips');
+const log = document.getElementById('log');
+
+let tabId = null;
+
+async function getArenaTab() {
+  const [tab] = await chrome.tabs.query({ active: true, currentWindow: true });
+  if (tab && tab.url && tab.url.startsWith('https://artificialanalysis.ai/text-to-speech/arena')) {
+    return tab;
+  }
+  return null;
+}
+
+function updateUI(state) {
+  if (state.ws_connected) {
+    statusText.textContent = 'Connected';
+    dot.className = 'dot on';
+  } else {
+    statusText.textContent = 'Disconnected';
+    dot.className = 'dot off';
+  }
+  serverStatus.textContent = state.ws_connected ? 'Online' : 'Offline';
+}
+
+function addLog(msg) {
+  const time = new Date().toLocaleTimeString();
+  log.textContent = `[${time}] ${msg}\n` + log.textContent;
+}
+
+async function refreshState() {
+  const tab = await getArenaTab();
+
+  chrome.runtime.sendMessage({ type: 'get_state' }, (state) => {
+    updateUI(state);
+    if (state.today_votes !== undefined) {
+      todayVotes.textContent = `${state.today_votes} / ${state.daily_limit || 800}`;
+      todaySkips.textContent = state.today_skips || 0;
+    }
+  });
+
+  if (!tab) return;
+  tabId = tab.id;
+
+  chrome.tabs.sendMessage(tabId, { type: 'get_status' }).then((contentState) => {
+    if (contentState) {
+      btnStart.disabled = contentState.enabled;
+      btnStop.disabled = !contentState.enabled;
+    }
+  }).catch(() => {});
+}
+
+btnStart.addEventListener('click', async () => {
+  const tab = await getArenaTab();
+  if (!tab) {
+    addLog('Arena page not found. Open it first.');
+    return;
+  }
+  chrome.tabs.sendMessage(tab.id, { type: 'start' }).then((res) => {
+    if (res) {
+      btnStart.disabled = true;
+      btnStop.disabled = false;
+      addLog('Voting started');
+    }
+  }).catch(() => {
+    addLog('Failed to start. Reload the arena page.');
+  });
+});
+
+btnStop.addEventListener('click', async () => {
+  const tab = await getArenaTab();
+  if (!tab) return;
+  chrome.tabs.sendMessage(tab.id, { type: 'stop' }).then((res) => {
+    if (res) {
+      btnStart.disabled = false;
+      btnStop.disabled = true;
+      addLog('Voting stopped');
+    }
+  }).catch(() => {});
+});
+
+// Listen for updates from background
+chrome.runtime.onMessage.addListener((message) => {
+  if (message.type === 'from_server') {
+    const payload = message.payload;
+    if (payload.stats) {
+      todayVotes.textContent = `${payload.stats.today_votes} / ${payload.stats.daily_limit}`;
+      todaySkips.textContent = payload.stats.today_skips;
+    }
+  }
+  if (message.type === 'ws_state') {
+    updateUI({ ws_connected: message.status === 'connected' });
+  }
+});
+
+refreshState();
+setInterval(refreshState, 5000);

requirements.txt

+resemblyzer
+numpy
+scipy
+torch
+torchaudio
+websockets

server/config.py

+"""All tunable parameters for the Luna TTS ranking bot."""
+
+# ---- Matching ----
+OUR_THRESHOLD = 0.80          # Rule 1-3: min cosine similarity to claim "our speaker"
+COMPETITOR_THRESHOLD = 0.80   # Rule 4: competitor match threshold
+
+# ---- Voting strategy ----
+BIG_VENDOR_VOTE_PROBABILITY = 75  # Rule 5: percentage (0-100), roll <= this → vote
+DAILY_VOTE_LIMIT = 800            # Rule 6: stop after this many votes today
+
+# ---- Anti-detection timing ----
+VOTE_PERIOD_SECONDS = 600     # Vote for 10 min
+REST_MIN_SECONDS = 60         # Then rest 1-3 min (random)
+REST_MAX_SECONDS = 180
+VOTE_INTERVAL_MIN_SECONDS = 2 # Gap between votes
+VOTE_INTERVAL_MAX_SECONDS = 8
+CLICK_DELAY_MIN_MS = 100      # Extra delay before simulating keypress
+CLICK_DELAY_MAX_MS = 500
+
+# ---- Server ----
+WS_HOST = "localhost"
+WS_PORT = 8765
+
+# ---- Paths (relative to server.py location) ----
+REFERENCE_DIRS = {
+    "our": "reference/standard",
+    "competitor": "reference/competitors",
+    "big_vendor": "reference/big_vendors",
+}

server/install_host.sh

+#!/bin/bash
+# Install Chrome Native Messaging host manifest (Mac)
+
+EXT_ID="$1"
+if [ -z "$EXT_ID" ]; then
+  echo "Usage: ./install_host.sh <extension-id>"
+  echo "Find your extension ID at chrome://extensions/"
+  echo "Example: ./install_host.sh abcdefghijklmnop"
+  exit 1
+fi
+
+HOST_DIR="$(cd "$(dirname "$0")" && pwd)"
+
+# Compile wrapper binary
+if [ ! -x "$HOST_DIR/run_host" ]; then
+  echo "Compiling native host wrapper..."
+  if ! command -v gcc &>/dev/null; then
+    echo ""
+    echo "========================================"
+    echo "ERROR: gcc 未安装"
+    echo "请先安装 Xcode Command Line Tools："
+    echo "  xcode-select --install"
+    echo "========================================"
+    echo ""
+    exit 1
+  fi
+  gcc -o "$HOST_DIR/run_host" "$HOST_DIR/runner.c" || {
+    echo "编译失败，请确认 gcc 可用"
+    exit 1
+  }
+  echo "  编译完成: $HOST_DIR/run_host"
+fi
+
+# Install manifest to all known Chromium browser directories
+INSTALLED=0
+
+install_to() {
+  local base="$1"
+  [ ! -d "$base" ] && return
+  local nm_dir="$base/NativeMessagingHosts"
+  mkdir -p "$nm_dir"
+  cat > "$nm_dir/com.luna.tts.json" << JSONEOF
+{
+  "name": "com.luna.tts",
+  "description": "Luna TTS Ranking Bot - Native Messaging Host",
+  "path": "$HOST_DIR/run_host",
+  "type": "stdio",
+  "allowed_origins": ["chrome-extension://$EXT_ID/"]
+}
+JSONEOF
+  echo "  $nm_dir/com.luna.tts.json"
+  INSTALLED=$((INSTALLED + 1))
+}
+
+for b in "Google/Chrome" "Chromium" "Quark" "Microsoft Edge"; do
+  install_to "$HOME/Library/Application Support/$b"
+done
+
+# AdsPower cache dirs
+ADSPOWER_BASE="$HOME/Library/Application Support/adspower_global/cwd_global/source/cache"
+if [ -d "$ADSPOWER_BASE" ]; then
+  for d in "$ADSPOWER_BASE"/*; do
+    [ -d "$d" ] && install_to "$d"
+  done
+fi
+
+if [ $INSTALLED -eq 0 ]; then
+  echo "WARNING: 没有找到浏览器目录，只装了以下基础路径："
+  NM_DIR="$HOME/Library/Application Support/Google/Chrome/NativeMessagingHosts"
+  mkdir -p "$NM_DIR"
+  cat > "$NM_DIR/com.luna.tts.json" << JSONEOF
+{
+  "name": "com.luna.tts",
+  "description": "Luna TTS Ranking Bot - Native Messaging Host",
+  "path": "$HOST_DIR/run_host",
+  "type": "stdio",
+  "allowed_origins": ["chrome-extension://$EXT_ID/"]
+}
+JSONEOF
+  echo "  $NM_DIR/com.luna.tts.json"
+fi
+
+echo ""
+echo "完成！Host: $HOST_DIR/run_host"
+echo "扩展 ID: $EXT_ID"
+echo ""
+echo "现在去 chrome://extensions/ 刷新扩展，然后打开打榜页面即可"

server/logs/crash.log

+
+========================================
+Traceback (most recent call last):
+  File "/Users/chengyubing/Documents/projects/hzjj/tts-ranking/server/run_host.py", line 12, in <module>
+    main()
+  File "/Users/chengyubing/Documents/projects/hzjj/tts-ranking/server/native_host.py", line 81, in main
+    print("[host] Ready, waiting for messages...", file=sys.stderr)
+BrokenPipeError: [Errno 32] Broken pipe
+