Create voices from natural language descriptions and synthesize studio-quality audio in 30 languages
Writing 'young woman, soft and sweet voice' without any reference audio generates a matching voice. VoxCPM2 is a text-to-speech system that generates continuous voice representations directly, bypassing discrete tokenization. This enables natural harmonics across 30 languages and precise voice cloning.
Changes in VoxCPM2, released in April 2026
Unlike previous versions, it supports 30 languages and outputs 48kHz studio-quality audio. The real-time factor (RTF) is approximately 0.3 on an NVIDIA RTX 4090, dropping to about 0.13 when using Nano-vLLM. It is released under the Apache-2.0 license, allowing commercial use.
Three methods for voice generation and cloning
VoxCPM2 generates voices in three different ways depending on user needs.
- Voice Design Creates new voices using only natural language descriptions without reference audio. You can specify gender, age, tone, emotion, and speed in text to generate the corresponding audio.
- Controllable Cloning Provides a short reference audio clip to maintain the original timbre while adjusting emotion, speed, and expression via control instructions.
- High-Quality Cloning Supplies both reference audio and its text transcript, allowing the model to naturally continue the reference voice while reproducing all details, including timbre, rhythm, emotion, and style.
Installation and execution
Install using the pip install voxcpm command in an environment with Python 3.10 or higher (below 3.13), PyTorch 2.5.0 or higher, and CUDA 12.0 or higher. Load the model via the Python API using VoxCPM.from_pretrained, then pass text and options to the generate method to create audio files. To access the web demo, start a local server with python app.py and open it in a browser.
Deployment environments and fine-tuning
Use Nano-vLLM-VoxCPM for high-throughput server deployments and vLLM-Omni for multi-tenant production environments. vLLM-Omni provides an OpenAI-compatible API, allowing direct calls from existing clients. To run on edge devices without Python, use llama.cpp-omni to execute GGUF format weights on CPU, Metal, CUDA, or Vulkan. Additionally, LoRA fine-tuning for specific speakers or domains is possible with just 5 to 10 minutes of audio data.
Considerations before use
VoxCPM2 generates highly realistic synthetic audio, so using it for impersonation, fraud, or spreading misinformation is strictly prohibited. It is recommended to clearly indicate that content is AI-generated. Voice design and controllable cloning may yield different results per run, so you may need to attempt generation 1 to 3 times to obtain the desired voice. While 30 languages are officially supported, languages not on the list require direct testing or fine-tuning with your own data. Thorough testing and safety evaluation for your specific use case are required before production deployment.
