ComfyUI-SoulX-Podcast is a custom node plugin for ComfyUI that packages the core functionality of SoulX-Podcast (long-form, multi-speaker, multi-dialect podcast voice generation) into a visual node workflow.

✨ Key Features

🎙️ Two-Person Podcast Generation: Supports dialogue generation between two speakers
🌍 Multi-Dialect Support: Supports multiple Chinese dialects (requires dialect model)
📝 Flexible Dialogue Scripts: Define dialogues through simple script format
🎵 Prompt Audio Driven: Clone speaker voice characteristics using reference audio (Suno)
🔄 Long-Form Generation: Supports generation of long-form podcast content
🎛️ Visual Workflow: Complete the entire generation process through node connections in ComfyUI

📦 Installation & Requirements

Python Dependencies

Ensure your ComfyUI environment has the following key dependencies installed:

s3tokenizer
diffusers
torch (requires CUDA support)
transformers
onnxruntime (or onnxruntime-gpu)
einops
librosa
scipy

⚠️ Important Note: This project requires transformers==4.57.1. Please install with caution as other versions may cause compatibility issues.

Model Preparation

Model files must be placed in ComfyUI's standard model directory:

ComfyUI/models/TTS/[model_name]/

Directory Structure Example:

ComfyUI/
  └── models/
      └── TTS/
          └── SoulX-Podcast-1.7B/
              ├── soulxpodcast_config.json
              ├── flow.pt
              ├── hift.pt
              ├── campplus.onnx
              └── [LLM model files...]

Model Type Description:

Standard Model (e.g., SoulX-Podcast-1.7B): Suitable for standard Mandarin podcast generation
Dialect Model (e.g., SoulX-Podcast-1.7B-dialect): Supports multiple Chinese dialects, such as Henan dialect, Sichuan dialect, Cantonese, etc.

Important Note: To use dialect features, please ensure you load the SoulX-Podcast-1.7B-dialect model. Select the corresponding dialect model in the model_name parameter of the SoulX Podcast Loader node.

🚀 Quick Start

Basic Workflow

SoulX Podcast Loader - Load models
Load Audio - Prepare prompt audio for each speaker (using ComfyUI's "Load Audio" node)
SoulX Podcast Input Parser - Parse inputs and dialogue script
SoulX Podcast Generate - Generate podcast audio
Preview/Save Audio - Use ComfyUI's audio save node

Workflow Example

Here is a complete ComfyUI workflow example demonstrating how to use these nodes:

Import Workflow:

You can directly import the workflow JSON file in ComfyUI: example/example_workflow.json

This example includes:

Model loading configuration
Two-speaker prompt audio (male and female voices)
Complete dialogue script example
All necessary node connections

📖 Node Description

Node 1: SoulX Podcast Loader

Function: Loads all required models and tokenizers in one go.

Input Parameters

Parameter	Type	Default	Description
model_name	Dropdown	First available model	Select the model to use from `models/TTS/` directory
llm_engine	Selection	`hf`	LLM inference engine: `hf` (HuggingFace) or `vllm`
fp16_flow	Boolean	`False`	Whether to use FP16 precision for Flow model
seed	Integer	`1988`	Random seed, range: 0 to 4294967295

Output

SOULX_MODEL: Object containing all loaded models

Node 2: SoulX Podcast Input Parser

Function: Processes all input data (audio, text, dialogue script) and preprocesses it into a format usable by the model. Supports two-person dialogue (S1 and S2).

Required Inputs

Parameter	Type	Description
soulx_model	SOULX_MODEL	Model object from SoulX Podcast Loader node
input_mode	Selection	Input mode: `simple` (simple mode) or `json` (JSON mode)

Simple Mode Inputs (input_mode = "simple")

Parameter	Type	Description
S1_prompt_audio	AUDIO	Speaker 1 (S1) prompt audio for extracting voice characteristics
S2_prompt_audio	AUDIO	Speaker 2 (S2) prompt audio (optional, for two-person dialogue)
dialogue_script	Multi-line text	Dialogue script defining the entire podcast dialogue Format: `[S1] First sentence\n[S2] Second sentence` The system automatically extracts the first sentence from each speaker as prompt text

Output

PODCAST_INPUT: Object containing all preprocessed data

Processing Flow:

Parse dialogue script, extract text and speaker identifiers for each sentence
Extract speaker embedding features from prompt audio
Extract mel spectrogram features from audio
Convert text to token IDs
Package all preprocessed data

Node 3: SoulX Podcast Generate

Function: Executes the core inference loop to generate the final podcast audio.

Required Inputs

Parameter	Type	Description
soulx_model	SOULX_MODEL	Model object from Loader node
podcast_input	PODCAST_INPUT	Preprocessed data from Input Parser node
seed	Integer	Random seed
temperature	Float	LLM sampling temperature (default: 0.6, range: 0.1-2.0)
repetition_penalty	Float	Repetition penalty coefficient (default: 1.25, range: 1.0-2.0)

Optional Inputs

Parameter	Type	Default	Description
top_k	Integer	100	Top-K sampling (range: 1-200)
top_p	Float	0.9	Nucleus sampling (range: 0.1-1.0)
min_tokens	Integer	8	Minimum tokens per sentence (range: 1-100)
max_tokens	Integer	3000	Maximum tokens per sentence (range: 100-5000)

Output

AUDIO: Generated complete podcast audio (24kHz, ComfyUI AUDIO format compliant)

⚙️ Parameter Tuning Suggestions

Audio Quality Optimization

Prompt Audio Quality: Use clear, noise-free prompt audio, 5-15 seconds duration is optimal
fp16_flow: If you have sufficient VRAM, it's recommended to disable this for best audio quality

Generation Stability

temperature: 0.5-0.7 for dialogues, 0.7-0.9 for creative content
repetition_penalty: 1.15-1.35 is a good range
top_k / top_p: Keep defaults usually

Performance Optimization

llm_engine: Use vllm if available and you have sufficient VRAM for faster speed
fp16_flow: Enable to reduce VRAM usage
max_tokens: Adjust according to actual needs, don't set too large

🔧 FAQ

Q1: Model Loading Failed

Problem: Error message "Model path does not exist" or "No models found"

Solution:

Ensure model files are placed in ComfyUI/models/TTS/[model_name]/ directory
Check that the model directory contains all required files

Q2: Unstable Voice Characteristics in Generated Audio

Solution:

Use longer, clearer prompt audio (recommended about 10 seconds)
Ensure prompt audio quality is good with no obvious noise

Q3: Slow Generation Speed

Solution:

Use vllm engine if supported
Enable fp16_flow to reduce VRAM usage
Reduce max_tokens value

Q4: Dialogue Script Format Error

Wrong Example:

S1 Hello  # ❌ Missing brackets

Correct Format:

[S1] Hello  # ✅ Correct
[S2] Hello  # ✅ Correct

📚 Technical Architecture

Generation Pipeline

Text Input → LLM → Semantic Tokens → Flow → Mel Spectrogram → HiFT → Audio Waveform

Core Models

LLM: Qwen3-1.7B (Language Model)
Flow: CausalMaskedDiffWithXvec (Acoustic Model)
HiFT: HiFTGenerator (Vocoder)
Suno Extraction: campplus.onnx (Speaker Embedding Extraction)

📄 License

This project is developed based on the SoulX-Podcast project. Please follow the license requirements of the original project.

🙏 Acknowledgments

SoulX-Podcast - Original project
ComfyUI Community

📧 Support

For questions or suggestions, please submit them in the project Issues.