Qwen3-TTS Apple Silicon

SkillMedia

Local Qwen3-TTS on Apple Silicon using MLX. Use when generating Chinese speech, testing Qwen3-TTS voices, selecting built-in speakers, applying style instructions, cloning a voice from reference audio, or designing a new voice from text descriptions on this Mac. Best for local/offline TTS, Apple Silicon/MLX validation, prompt-to-speech generation, and reusable dubbing workflows.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Qwen3-TTS Apple Silicon skill

What this skill tells your AI

The instructions your AI receives, as published by qianleigood/crawclaw in skills-optional/qwen3-tts-apple-silicon/SKILL.md and read by ahel’s review.

在这台 Apple Silicon Mac 上,优先通过 scripts/qwen3_tts.py 使用本地 Qwen3-TTS。

选择哪条链路

  • synth:已知 speaker + instruction 的普通合成
  • clone:给参考音频 + 准确转写,做语音克隆
  • voice-design:只给文字描述,直接设计新音色
  • list-speakers:查看当前预置 speaker

快速开始

cd /Users/qianleilei/.crawclaw/workspace/skills/qwen3-tts-apple-silicon
bash scripts/setup_env.sh
python scripts/qwen3_tts.py list-speakers

普通合成:

python scripts/qwen3_tts.py synth \
  --text "今天我们先验证一条更稳妥的本地配音方案。" \
  --speaker serena \
  --instruction "用自然、清晰、平稳的中文女声朗读。" \
  --file-prefix demo

语音克隆:

python scripts/qwen3_tts.py clone \
  --text "这是一条用于验证参考音频克隆效果的中文句子。" \
  --ref-audio /absolute/path/to/ref.wav \
  --ref-text "这是一条用于验证参考音频克隆效果的中文句子。" \
  --file-prefix clone_demo

音色设计:

python scripts/qwen3_tts.py voice-design \
  --text "今天测试一下音色设计能力。" \
  --instruction "一个自然、温和、清晰的中文女声,成熟一些,语速平稳,适合讲解和说明。" \
  --file-prefix designed

默认策略

  • synth 默认模型:mlx-community/Qwen3-TTS-12Hz-0.6B-CustomVoice-8bit
  • clone 默认模型:mlx-community/Qwen3-TTS-12Hz-0.6B-Base-8bit
  • voice-design 默认模型:mlx-community/Qwen3-TTS-12Hz-1.7B-VoiceDesign-4bit
  • 默认语言:zh
  • 默认 speaker:serena
  • 默认输出目录:~/.cache/crawclaw/qwen3-tts-apple-silicon/outputs/
  • 先用 0.6B 验证链路;只有需要 VoiceDesign 时再拉起 1.7B

注意事项

  • 不要依赖 mlx_audio 的通用默认参数;对普通合成显式传 speakerlang=zh
  • clone 已绕过 generate_audio() 包装层,避免误导性的 Voice: af_heart 日志
  • 参考音频优先 5-10 秒、单人、安静、无混响,且 ref_text 尽量准确
  • 首次运行可能下载模型;未配置 Hugging Face token 时会更慢
  • 命令末尾会输出 JSON,包含 output_diraudio_files
  • 三条主命令都支持 --file-prefix

资源

  • scripts/setup_env.sh:创建 .venv 并安装依赖
  • scripts/qwen3_tts.py:统一 CLI,支持 list-speakers / synth / clone / voice-design
  • references/notes.md:实测数据、模型说明、af_heart 排查结论、warning 记录

Signals

GitHub stars
30
Forks
1
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
qwen3-tts-apple-silicon
Source
github.com/qianleigood/crawclaw