128 TTS Voice Lines into a Red Alert Engine: Qwen-TTS + Neil Batch Synthesis

Date: 2026-08-07

1. Goal

Give the Red Alert engine a reproducible bilingual commander voice: 64 EVA events × two languages = 128 WAVs, all batch-synthesized by TTS and injected into the browser’s OPFS. The principles: code committed, keys never on disk, regenerate anytime.

2. Why Qwen-TTS’s Neil

We surveyed voices across three Aliyun TTS APIs:

CosyVoice / Qwen-Audio-TTS: system voices are all longan/loong series; downloaded the 1000+ base-voice Excel files and searched — no Neil;
MiniMax: system voices use male-qn-* / female-* IDs, no Neil;
Qwen-TTS: has Neil (voice name “阿闻”, male news anchor, 10 languages incl. zh/en), supported by qwen3-tts-instruct-flash.

Neil’s news-anchor voice suits “commander announcements”, and instructions control pacing/emotion — one instruction per language.

3. Batch Generation Script

scripts/tts-generate.mts (TypeScript, tsc-checked):

const MODEL = "qwen3-tts-instruct-flash";
const VOICE = "Neil";
const INSTRUCTIONS = {
  en: "News anchor style, steady pace, calm and neutral emotion.",
  "zh-CN": "新闻播报风格，语速适中，吐字清晰，情感沉稳中性。",
};

Flow: read design/voice/{lang}/event_voice_mapping.json → call the DashScope API per line → download wav into audio/voice/{lang}/EVA_*.wav. Supports --lang / --event / --force for incremental regeneration.

4. Keys Never Touch Disk

API key comes only from the environment (process.env.DASHSCOPE_API_KEY); no hardcoding in the script;
.gitignore gained .env / .env.* (keeping .env.example) so git add . can’t commit secrets;
Request headers use a runtime Bearer ${API_KEY}; logs never echo the key.

Verified: searching all git-tracked files for the real key prefix → zero matches.

5. How Voice Reaches the Engine

Output: audio/voice/{en,zh-CN}/EVA_1MinuteRemaining.wav etc., 128 files;
The engine’s EVA announcer plays via eva.ini [DialogList] (Allied=cevaXXX) → vanilla wav from mix;
The engine consumption side of custom voice (event trigger → read design/voice mapping → play custom wav) is follow-up wiring; vanilla voice serves as fallback for now.

6. Injection Manifest Fix

gen-inject-page.mjs originally collected only files at the audio/ root — it did not recurse into audio/voice/{lang}/, so new voice lines were never injected. Changed to recursive collection (collectFiles walks subdirectories), and filtered .DS_Store (macOS junk files caused injection 404s).

7. Conclusion

The full pipeline for 128 TTS lines: voice research (Neil) → batch script (.mts + env key) → generate wav → inject (recursive manifest) → engine wiring (pending). Core principles: reproducible code, keys never on disk, output regenerable anytime — this skeleton also serves future voice iteration and engine wiring.

This article is licensed under Mulan Open Works License Attribution-PatentLicensed, Version 1 (MulanOWL BY-PL v1), or alternatively CC0-1.0 (public domain dedication). Copyright reserved. Attribution required. For enterprise compliance, please retain the original license notice. Original version: https://k4m7v2pz.github.io/en/blog/games/openyrweb-tts-voice-batch.html

Keyboard shortcuts

K4m7v2pz's Blog