VoxCPM

VoxCPM is an open-source text-to-speech system from OpenBMB, with the website copyright held by Modelbest Technology Co., Ltd. The current release, VoxCPM2, is a 2B-parameter model that generates multilingual speech, designs voices from text descriptions, and clones voices from short reference clips, outputting 48kHz audio. The code and weights are released under the Apache-2.0 license.

What It Is

VoxCPM is a tokenizer-free TTS model. Instead of discrete speech tokens, it uses a diffusion autoregressive architecture to generate continuous speech representations. VoxCPM2 is built on a MiniCPM-4 backbone, and the README says it was trained on over 2 million hours of multilingual speech data. It accepts text in any of 30 supported languages without a language tag, plus several Chinese dialects.

Voice Creation and Cloning

There are three main modes. Voice Design creates a new voice from a natural-language description placed in parentheses at the start of the text, with no reference audio. Controllable Cloning copies the timbre of a short reference clip while allowing style guidance for emotion, pace, and expression. Ultimate Cloning takes a reference clip plus its transcript and continues from it to preserve timbre, rhythm, emotion, and style. The README notes that Voice Design and Controllable Cloning results can vary between runs, so several generations may be needed.

Running and Deploying It

The model installs with pip and is used through a Python API (including streaming), a CLI with design, clone, and batch commands, and a local web demo. For serving, the README points to Nano-vLLM-VoxCPM and vLLM-Omni, which exposes an OpenAI-compatible speech endpoint, and to llama.cpp-omni for on-device GGUF inference on CPU, Metal, CUDA, and Vulkan. The README lists Python 3.10 to below 3.13, PyTorch 2.5.0 or later, and CUDA 12.0 or later as requirements, and about 8 GB of VRAM for VoxCPM2.

Fine-tuning and Ecosystem

Full fine-tuning and LoRA fine-tuning are supported, and the README says 5 to 10 minutes of audio can adapt the model to a speaker, language, or domain. A WebUI is provided for training and inference. Community projects such as ComfyUI nodes, ONNX export, and Rust and C++ ports are listed but are not officially maintained by OpenBMB.

Current Status

The README lists VoxCPM2 as the latest release, released in April 2026, with VoxCPM1.5 as stable and VoxCPM-0.5B as legacy. The latest GitHub release is v2.0.3, published 2026-05-11. The README forbids using voice cloning for impersonation, fraud, or disinformation.