🎙️ VoxCPM WebUI: Voice Cloning & Fine-Tuning

A comprehensive WebUI for fine-tuning and interacting with VoxCPM models. This application leverages Efficient LoRA (Low-Rank Adaptation) capabilities to enable high-quality voice cloning.

---

🔄 Application Workflow

1. Data Preparation: Upload/record audio and generate transcriptions using Faster-Whisper.
2. Fine-Tuning: Configure LoRA or Full Fine-Tuning parameters and train your voice adapter.
3. Inference: Generate speech using the base model combined with your trained weights.

• Persistent Torch & Triton cache: Integration of triton-windows and a custom kernel caching system in models/.cache, enabling the full power of torch.compile for inference speed-up.
• Building the persistent cache for the first time, might take a up to 5 minutes. This is a one time process. Once cached, Subsequent generations will be significantly faster.

---

⚙️ System Requirements & Hardware

💻 Software Dependencies

• Python: 3.10 – 3.11 (Recommended for stability during training).
• PyTorch: 2.5.0+
• CUDA: 12.0+
• Format Support: .wav is recommended.

🔌 Hardware Setup (VRAM Requirements)

Model	LoRA Training
VoxCPM 1.5 (750M)	~12 GB VRAM
VoxCPM 2.0 (2B)	~20 GB VRAM

---

📊 Dataset & Audio Specifications

🎯 Clip Requirements

• Format: .wav is highly recommended. Other formats supported by torchaudio also work.
• Duration: 3–30 seconds per clip is the "sweet spot."
  • Warning: Clips < 1s produce unstable results.
  • Warning: Very long clips increase VRAM usage and may be filtered by max_batch_tokens.
• Sample Rate: The dataloader resamples automatically. Your config sample_rate must match the AudioVAE encoder input:
  • VoxCPM 1.0: 16kHz
  • VoxCPM 1.5: 44.1kHz
  • VoxCPM 2.0: 16kHz (The encoder operates at 16kHz; the decoder outputs 48kHz).

✨ Preprocessing Tips

• Trim Trailing Silence: Keep silence to < 0.5 seconds. Excessive trailing silence is the leading cause of "infinite generation" issues after fine-tuning.
• Normalize Volume: Ensure consistent levels across all training samples.
• Clean Transcripts: Text must match audio exactly. Inaccurate transcripts degrade both cloning quality and text adherence.
• Remove Noise: The model is highly sensitive to background noise. Use clean, isolated voice recordings.

💻 Hardware Requirements

• Inference (Running the model):
  • Minimum: 8 GB VRAM
  • Recommended: 12 GB VRAM
• Training (LoRA):
  • Minimum: +12 GB VRAM (VoxCPM 1.5)
  • Recommended: +20 GB VRAM (VoxCPM 2.0)

---

Clone the repository:

git clone https://github.com/OpenBMB/VoxCPM.git

🛠️ Installation & Execution (Windows)

This project utilizes uv for lightning-fast dependency management.

Setup Steps

1. Run Installer: Double-click install.bat.
   • This installs uv via Winget (if not present).
   • Synchronizes the environment and installs all required libraries automatically.
2. Launch App: Double-click start.bat.
3. Access: Navigate to http://127.0.0.1:7860 in your web browser.

---

Inspired by FranckyB Voice Clone Studio

Based on VoxCPM2 by OpenBMB