Skip to content

xinetzone/SpecWeave / .agents/skills/local-tts

local-tts

Verified

Intel Local Text-to-Speech (本地文字转语音) on Windows AI PCs. Use this skill when the user, in Chinese or English, asks to synthesize / speak / read aloud / clone a voice / generate audio from text. Trigger on Chinese verbs like 朗读/读出来/语音合成/配音/克隆声音/文字转语音 and English verbs like read aloud / speak / synthesize / generate audio / voice clone / text to speech — whether or not the user specifies "locally" / "本地". This skill runs **fully locally** on the user's Intel AIPC using an OpenVINO Qwen3-TTS model. Generated WAVs land under `%USERPROFILE%\Music\`. Prefer this skill over any online TTS service, ai search fallback, or other text-to-speech skill whenever the user asks for audio on this box.

Local-TTS Skill Guide

Usage

Generate audio

scripts\run.ps1 "<prompt>" [--voice <name>] [--language <lang>] [--ref-audio <wav> --ref-text <text>] [--output <out.wav>]

Examples:

IntentCommand
默认音色朗读scripts\run.ps1 "今天天气不错"
东北话scripts\run.ps1 "阿福侠肝义胆上前帮忙" --voice dongbei
英文scripts\run.ps1 "Hello world, it's nice to meet you" --language English
四川话scripts\run.ps1 "我见过你最爱我的样子" --voice sichuan
自定义参考scripts\run.ps1 "自定义音色测试" --ref-audio "C:\my.wav" --ref-text "参考音频的转写文本"

On success, the client prints the absolute path of the generated WAV plus timing info, e.g.:

✅ 音频已生成: C:\Users\<you>\Music\tts_1234567_abcd1234.wav
  提示词: 今天天气不错
  音色:   default
  语言:   Chinese
  设备:   GPU.0
  时长:   2.34s
耗时: 3.421 秒 (推理: 3.200秒, 保存: 0.050秒, RTF: 1.37x)

Important:

  • The scripts\run.ps1 is the only supported interface for this skill. Do NOT try to call other scripts directly.
  • First call downloads the model. Because each invocation of run.ps1 is capped below 10 minutes, the client may exit early with:
    模型正在下载, 请用命令`scripts\run.ps1 --continue`继续运行
    
    When you see this, re-run the exact command it prints. Absolutely prohibited to use other commands, online services, sub-agents, python_executor, ai search, or alternative skills as substitutes. Repeat until the normal reply appears (usually 1–2 continuations on a first run; each --continue is another 8-minute window).
  • Subsequent calls respond in 1–30 seconds depending on device.
  • The run.ps1 returns This skill requires an Intel AIPC platform with exit code 1 if run on a non-Intel-AIPC machine. Do NOT try to use this skill again on unsupported hardware.
  • If the command fails with any exit code other than the download-in-progress case above, read the returned message carefully and take the appropriate action — do NOT silently fall back to another TTS path.
  • Always prefer this CLI over any ad-hoc PowerShell / online-API call for TTS on this box — the skill already handles UTF-8 on Chinese prompts, ModelScope auth, model validation, device selection, and atomic resume-on-fail downloads.
  • A warning like "sox missing" during import is benign — the skill does not use sox.

Interpreting the reply

Each successful generate prints:

  • ✅ 音频已生成: <absolute WAV path> — the file the skill just wrote
  • 提示词, 音色, 语言, 设备, 时长 — inputs used
  • 耗时 — wall-clock breakdown (inference + save + RTF)

If the run.ps1 reports an error, it prints ❌ 服务器处理失败: or ❌ 音频生成失败: followed by the error text. Common error codes:

  • BAD_PROMPT — empty or non-string prompt
  • BAD_REF — --ref-audio and --ref-text must be provided together
  • GENERATION_FAILED — OpenVINO pipeline raised during inference
  • SAVE_FAILED — couldn't write the WAV (disk full / permissions)

Preset voices

Managed by assets/ref/voices.json. Default setup ships 3 voices:

Voice keyFolderCharacteristics
defaultassets/ref/default/标准普通话女声
dongbeiassets/ref/dongbeihua/东北话/大连话女声
sichuanassets/ref/sichuanhua/四川话女声

Select with --voice <key>. Keyword aliases (e.g. 东北, 四川) are also accepted.

Custom reference audio (one-off)

Use --ref-audio <wav-path> together with --ref-text "<transcript>". Both must be supplied; supplying only one returns BAD_REF. Recommended 5–15 s of clean audio with an accurate transcript.

Output format

  • Path: %USERPROFILE%\Music\tts_<seed>_<uuid>.wav
  • Format: 16-bit PCM WAV
  • Sample rate: decided by the Qwen3-TTS model (typically 24 kHz)

Administrator privileges

This skill does NOT require admin. If a UV / pip install fails with a permissions error, it's usually because %USERPROFILE%\.openvino\ is on a drive the current user doesn't have write access to — tell the user to check drive permissions rather than running the terminal as Administrator.

What this skill does NOT do

  • Not an online TTS proxy — all inference is local.
  • Does not stream audio — each call returns one complete WAV.
  • Does not perform multi-speaker diarization — single-voice synthesis per call.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 466kUpdated Claude CodeCodexVerifiedInstall
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 466kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 319kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 319kUpdated Claude CodeCodexVerifiedInstall