VoxCPM

Most TTS systems turn speech into discrete tokens first; VoxCPM skips that step and generates continuous speech representations with an end-to-end diffusion autoregressive architecture. VoxCPM2, the current release from OpenBMB, is a 2B-parameter model on a MiniCPM-4 backbone that the project says was trained on more than 2 million hours of speech, covering 30 languages including Korean and Japanese, plus Chinese dialects such as Cantonese and Sichuanese, without a language tag. Voice Design creates a new voice from a text description of gender, age, tone or pace, while cloning works from a short reference clip, optionally with its transcript for closer reproduction. It takes 16kHz reference audio and outputs 48kHz, and the project reports a real-time factor around 0.3 on an RTX 4090, or about 0.13 with Nano-vLLM or vLLM-Omni. Weights and code are released under Apache-2.0, which allows commercial use.