This extension not only perfectly reproduces the core capabilities of Qwen3-TTS but extends it with a Smart Script Engine, Multi-Speaker Cinematic Engine, and Professional Audio Post-Processing. It covers everything from simple single-sentence synthesis to complex, full-cast audio dramas.

Video and Article coming Soon

📋 Changelog

2026-01-24: Core Architecture Upgrade
- New Nodes: Added RoleBank (Character Registry), AdvancedDialogue (Multi-speaker Engine), and ScriptProcessor.
- Smart Parsing: Added support for auto-parsing [Emotion Tags] and [pause:1.2s] timing instructions within text.
- Control: Added Seed control for deterministic generation, Output Mode (Merge/Split), and Audio-to-Text (ASR) capabilities.
2026-01-23: Sampling Controls
- Basic Features: Unlocked full sampling parameters for all generation nodes (top_p, top_k, temperature, repetition_penalty) to prevent audio degradation.

✨ Key Features

1. The Four Core Capabilities

🎭 Custom Voice (Presets): Access 9 high-quality internal presets immediately (e.g., Vivian, Uncle_Fu).
🎨 Voice Design (Text-to-Voice): Design unique voices using natural language prompts (e.g., "A deep, rasping voice of an old wizard").
🦜 Voice Clone (Zero-Shot): Clone a voice using just 3+ seconds of reference audio. Includes an X-Vector Mode for pure timbre extraction without reference text.
⚡ Pre-Compute (Latents): Decouples voice analysis from generation. Analyze the reference audio once, and generate infinite text rapidly without re-processing the source.

2. Advanced Cinematic Features (New)

🧠 Smart Sentiment Parsing: Automatically detects and applies bracketed emotion tags like [Happy], [Cold], or [Whisper].
⏱️ Precision Timing Control: Use [pause:1.2] inside your script to insert exact milliseconds of silence for pacing.
👥 Multi-Role Orchestration: Use the RoleBank to register multiple voices. The AdvancedDialogue node then automatically switches voices based on your script, generating a full conversation in one go.
💾 Asset Persistence: Save your cloned voice embeddings as .qwen3tts files to build a library of characters that can be reused anytime.

3. Engineering Optimizations

Deterministic Output: Full Seed support allows you to lock the random noise. If you like a specific take, you can reproduce it exactly.
Batch Inference: Supports multi-line input for parallel generation.
Dual Download Sources: Supports automatic switching between ModelScope (CN) and HuggingFace.
Robust Execution: Built-in VRAM management, forced garbage collection, and CPU fallback protection to prevent OOM crashes.

📦 Installation

1. Plugin Installation

Navigate to your ComfyUI custom_nodes directory:

cd ComfyUI/custom_nodes
git clone [https://github.com/wanaigc/ComfyUI-Qwen3-TTS.git](https://github.com/wanaigc/ComfyUI-Qwen3-TTS.git)
cd ComfyUI-Qwen3-TTS