FL VoxCPM

Text-to-speech nodes for ComfyUI powered by OpenBMB's VoxCPM model family. Features V1 and V2 models, voice cloning, voice design from text descriptions, LoRA fine-tuning with real-time training dashboard, and audio utility nodes.

Features

• VoxCPM V2 - 2B parameter model, 48kHz studio-quality audio, 30 languages
• VoxCPM V1.5 - 44.1kHz high-fidelity TTS with 800M parameters
• Voice Design - Create custom voices from natural language descriptions (V2)
• Voice Cloning - Clone any voice from a short audio reference
• Controllable Cloning - Clone a voice and modify its style/emotion (V2)
• Ultimate Cloning - Maximum fidelity cloning with reference + continuation audio (V2)
• LoRA Training - Fine-tune custom voices with real-time dashboard (loss chart, progress, validation audio)
• Auto Transcription - Built-in Whisper integration for generating reference text
• Audio Crop - Trim audio to specific time ranges

Nodes

Installation

ComfyUI Manager

Search for "FL VoxCPM" and install.

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI-FL-VoxCPM.git
cd ComfyUI-FL-VoxCPM
pip install -r requirements.txt

Quick Start

V2 Voice Design (No Reference Audio Needed)

1. Add FL VoxCPM V2 TTS node
2. Set mode to Voice Design
3. Enter voice description in control (e.g. "young woman, warm and gentle voice")
4. Enter target text, queue prompt

V2 Voice Cloning

1. Add FL VoxCPM V2 TTS + Load Audio nodes
2. Set mode to Voice Cloning
3. Connect reference audio to reference_audio
4. Enter target text, queue prompt

V1 Text-to-Speech

1. Add FL VoxCPM TTS node
2. Enter text, queue prompt
3. Model downloads automatically on first run

V1 Voice Cloning

1. Add FL VoxCPM TTS + Load Audio nodes
2. Connect reference audio to prompt_audio
3. Use FL VoxCPM Transcribe to generate transcript, or enter it manually in prompt_text
4. Enter target text, queue prompt

LoRA Training

1. Prepare folder with .wav/.mp3/.flac files + matching .txt transcripts
2. FL VoxCPM Dataset Maker -> FL VoxCPM Train Config (or V2) -> FL VoxCPM LoRA Trainer
3. Set validation_text to hear samples at each checkpoint
4. Monitor training in the real-time dashboard (loss chart, progress bar, audio samples)
5. LoRA saves to ComfyUI/models/loras/VoxCPM/

Models

Models auto-download from HuggingFace to ComfyUI/models/tts/VoxCPM/ on first use.

V2 Modes

Key Parameters

• cfg_value (2.0) - Guidance scale. Higher = closer to reference voice, lower = more natural
• inference_timesteps (10) - Diffusion steps. More = better quality, slower
• lora_name - Select trained LoRA from dropdown (auto-detects rank)
• control (V2 only) - Natural language voice description for design/style control

Dataset Format (Training)

audio_folder/
  sample1.wav
  sample1.txt    # contains transcript of sample1.wav
  sample2.mp3
  sample2.txt
  ...

Supported audio formats: .wav, .mp3, .flac

Requirements

• Python 3.9+
• 16GB RAM minimum (32GB+ recommended for training)
• NVIDIA GPU with 12GB+ VRAM recommended (CPU and Mac MPS supported for V1 inference)
• V2 model requires ~10GB VRAM for inference, ~20GB for training

License

Apache-2.0 - Based on OpenBMB/VoxCPM