This is a test wrapper // code generated by claude

Qwen3-TTS ComfyUI Nodes

Fast and efficient Text-to-Speech nodes for ComfyUI based on Qwen3-TTS.

Features

• 🎙️ Custom Voice Generation: 9 premium voices with emotion/style control
• 🎨 Voice Design: Create unique voices from text descriptions
• 👥 Voice Cloning: Clone any voice from 3+ seconds of audio
• 🌍 Multilingual: Supports 10 languages (Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian)
• ⚡ Fast: Optimized with model caching and flash attention
• 📦 Batch Processing: Generate multiple audio clips efficiently

Installation

Method 1: Via ComfyUI Manager (Recommended)

1. Open ComfyUI Manager
2. Search for "Qwen3-TTS"
3. Click Install
4. Restart ComfyUI

Method 2: Manual Installation

1. Navigate to your ComfyUI custom nodes directory:

cd ComfyUI/custom_nodes/

2. Clone this repository:

git clone https://github.com/yourusername/ComfyUI-Qwen3-TTS

3. Install dependencies:

cd ComfyUI-Qwen3-TTS
pip install qwen-tts soundfile torch

4. (Optional) Install FlashAttention for better performance:

pip install flash-attn --no-build-isolation

5. Restart ComfyUI

Nodes

1. Qwen3 TTS Model Loader

Loads and caches Qwen3-TTS models.

Inputs:

• model_name: Choose from available models (1.7B or 0.6B variants)
• device: GPU or CPU selection
• dtype: Precision (bfloat16, float16, float32)
• use_flash_attn: Enable FlashAttention 2 for speed

Outputs:

• QWEN_TTS_MODEL: Loaded model instance

2. Qwen3 TTS Custom Voice

Generate speech using one of 9 premium voices.

Available Speakers:

• Vivian: Bright, slightly edgy young female (Chinese)
• Serena: Warm, gentle young female (Chinese)
• Uncle_Fu: Seasoned male with low, mellow timbre (Chinese)
• Dylan: Youthful Beijing male (Chinese - Beijing Dialect)
• Eric: Lively Chengdu male (Chinese - Sichuan Dialect)
• Ryan: Dynamic male with strong rhythm (English)
• Aiden: Sunny American male (English)
• Ono_Anna: Playful Japanese female (Japanese)
• Sohee: Warm Korean female (Korean)

Inputs:

• model: Qwen3-TTS model
• text: Text to synthesize
• speaker: Voice selection
• language: Target language (or "Auto")
• instruct (optional): Style/emotion instruction (e.g., "Very happy", "Angry tone")

3. Qwen3 TTS Voice Design

Create custom voices from text descriptions.

Inputs:

• model: Qwen3-TTS VoiceDesign model
• text: Text to synthesize
• voice_description: Describe the desired voice (e.g., "A deep, authoritative male voice with British accent")
• language: Target language

Example Descriptions:

• "A warm, gentle young female voice with clear pronunciation"
• "Deep masculine voice with slight rasp, confident and commanding"
• "Cheerful teenage girl voice, energetic and playful"
• "Elderly male voice, wise and calm, with slight tremor"

4. Qwen3 TTS Voice Clone

Clone a voice from reference audio.

Inputs:

• model: Qwen3-TTS Base model
• text: Text to synthesize
• ref_audio: Reference audio (ComfyUI AUDIO format)
• ref_text: Transcript of reference audio
• language: Target language
• x_vector_only (optional): Use only speaker embedding (faster but lower quality)

5. Qwen3 TTS Voice Clone (File)

Clone voice from file path or URL.

Inputs:

• ref_audio_path: Local file path or URL
• Other inputs same as Voice Clone

Example URLs:

https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-TTS-Repo/clone.wav

6. Qwen3 TTS Create Clone Prompt

Create reusable clone prompt for batch generation (optimization).

Inputs:

• model: Qwen3-TTS Base model
• ref_audio: Reference audio
• ref_text: Transcript

Outputs:

• QWEN_CLONE_PROMPT: Reusable prompt for faster generation

7. Qwen3 TTS Clone with Prompt

Generate speech using pre-created clone prompt.

Inputs:

• model: Qwen3-TTS Base model
• text: Text to synthesize
• clone_prompt: From "Create Clone Prompt" node
• language: Target language

8. Qwen3 TTS Batch Generate

Generate multiple audio clips from a list of texts.

Inputs:

• model: Qwen3-TTS model
• text_list: Multiple texts separated by delimiter
• separator: Text delimiter (default: "\n")
• speaker (optional): For custom voice mode
• language: Target language
• clone_prompt (optional): For voice cloning mode

Workflow Examples

Basic Custom Voice Generation

Model Loader → Custom Voice → Preview Audio → Save Audio

Voice Design

Model Loader (VoiceDesign) → Voice Design → Preview Audio

Voice Cloning

Load Audio → Model Loader (Base) → Voice Clone → Preview Audio

Optimized Batch Cloning

Load Audio → Model Loader (Base) → Create Clone Prompt → Clone with Prompt (×3) → Batch → Save

Performance Tips

1. Use Model Caching: The Model Loader caches models, so load once and reuse
2. Enable FlashAttention: Set use_flash_attn=True for 2-3x speed improvement
3. Use bfloat16: Best balance of speed and quality
4. Batch Processing: Use the Batch node or Clone with Prompt for multiple generations
5. GPU Selection: Use cuda:0 for best performance

Models

Models auto-download on first use from Hugging Face. You can also pre-download:

# Via ModelScope (faster in China)
pip install modelscope
modelscope download --model Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --local_dir ./models/Qwen3-TTS-12Hz-1.7B-CustomVoice

# Via Hugging Face
pip install huggingface_hub[cli]
huggingface-cli download Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --local-dir ./models/Qwen3-TTS-12Hz-1.7B-CustomVoice

Available Models:

• Qwen3-TTS-12Hz-1.7B-CustomVoice (9 premium voices)
• Qwen3-TTS-12Hz-1.7B-VoiceDesign (voice design)
• Qwen3-TTS-12Hz-1.7B-Base (voice cloning)
• Qwen3-TTS-12Hz-0.6B-CustomVoice (smaller, faster)
• Qwen3-TTS-12Hz-0.6B-Base (smaller, faster)

Requirements

• Python 3.8+
• PyTorch 2.0+
• qwen-tts
• soundfile
• CUDA-capable GPU recommended (CPU works but slower)

Troubleshooting

Model download fails:

• Set environment variable: export HF_ENDPOINT=https://hf-mirror.com (China users)
• Or manually download models as shown above

Flash Attention not available:

• Install with: pip install flash-attn --no-build-isolation
• Or disable in Model Loader settings

Out of memory:

• Use 0.6B model instead of 1.7B
• Use float16 or reduce batch size
• Close other applications

Poor voice quality:

• Use bfloat16 instead of float16
• Ensure reference audio is clean (for cloning)
• Provide accurate transcript for reference audio

License

Apache 2.0 (same as Qwen3-TTS)

Credits

Based on Qwen3-TTS by Alibaba Cloud Qwen Team.

Support

• Report issues: GitHub Issues
• Qwen3-TTS docs: Official Repository