What it is
VoxCPM2 is a 2B-parameter speech synthesis model that generates continuous audio directly, without discrete speech tokens, using a diffusion autoregressive design. It reads text in 30 languages and several Chinese dialects, outputs 48kHz audio, can invent a voice from a written description, and can copy a speaker's voice from just a few seconds of sample audio. Developers and audio creators load it through a Python package, CLI, or web demo, stream output in near real time, and fine-tune it with SFT or LoRA.