Qwen3-TTS VoiceDesign

Text → Speech with natural language voice descriptions + seed-based timbre fixation.

Quick Start

CODEBLOCK0

Environment Variables

All config via env vars — text is the only required argument:

Server reads from .env file in its directory. Client scripts read from shell env.

Voice Description Example

CODEBLOCK1

Tip: Once you've found your perfect voice (description + seed), set them as server defaults in .env. Then client calls only need to pass text.

API

OpenAI-Compatible

CODEBLOCK2

Custom (seed + instruct override)

CODEBLOCK3

GET (quick test)

CODEBLOCK4

Seed Mechanics

Same (description + seed) → same timbre. Different seeds → completely different voices.

⚠️ Seeds are purely random — seed 42 and 43 can sound completely different. Finding a voice = opening blind boxes.

Workflow: fix description → batch 30-40 seeds → listen → shortlist 2-3 → compare across scenarios → pick.

Deploy Your Own

CODEBLOCK5

Setup installs: qwen-tts, soundfile, pydub, uvicorn, fastapi, torch (CUDA).
Downloads VoiceDesign model (~3.5GB) via ModelScope (China) or HuggingFace.

Requirements: CUDA GPU with 4GB+ VRAM, Python 3.10+, ~4GB disk.

Scripts

OpenClaw Integration

In openclaw.json:
CODEBLOCK6

Server Management

CODEBLOCK7

Troubleshooting

• - Connection refused → Server not running; start it
• 30s+ first request → Cold start (model loading ~60s); subsequent requests 10-15s
• Behind proxy → Set NO_PROXY=<server_ip> on client side
• Windows firewall → INLINECODE33
• No flash-attn on Windows → Expected; falls back to PyTorch SDPA (slower but works)
• PowerShell corrupts Chinese → Edit .env/config via Python or SCP, not PowerShell INLINECODE35
• Process dies on SSH disconnect → Use scheduled task (Windows) or systemd (Linux) instead of foreground

Voice Design Tips

Describe like casting a voice actor:

• - Age/gender: "18岁女大学生" / "30岁男性播音员"
• Texture: "柔和温暖" / "清脆明亮" / "低沉磁性"
• Emotion: "轻柔细腻" / "活泼开朗"
• Accent: "南方口音软糯" / "台湾腔" / "东北大碴子味"
• Metaphor: "像棉花糖" / "像播音主持" (helps the model capture feeling)

⚠️ Timbre ≠ description. Description controls style/emotion; seed controls timbre. Don't put personality traits ("灵动俏皮") in description — that's the seed's job.