VoxCPM2

VoxCPM2 is OpenBMB's speech synthesiser, published on Hugging Face on 3 April 2026. It carries 2,290,004,544 parameters and was trained on more than two million hours of multilingual speech, and its architecture is unusual for the field: it is tokenizer-free and diffusion autoregressive, meaning it does not quantise speech into discrete audio tokens before generating it. The capability that sets it apart is voice design. Most cloning systems need a sample of the voice they are to imitate; VoxCPM2 will also build a new one from a description in plain language — gender, age, tone, emotion, pace — with no reference audio at all. Cloning proper is offered in two grades: a controllable mode that takes a short clip and lets style guidance steer emotion and pace while keeping the timbre, and a higher-fidelity mode that takes the reference clip together with its transcript and continues from it. Output is 48 kHz — studio sample rate — produced from 16 kHz reference material by super-resolution built into the AudioVAE V2 decoder, so no external upsampler is needed. OpenBMB reports a real-time factor of about 0.3 on an RTX 4090, falling to roughly 0.13 with the Nano-VLLM accelerator, which puts real-time streaming within reach of one consumer card. The model also infers prosody from the content of the text rather than requiring it to be marked up. Thirty languages are supported without a language tag, Polish among them, alongside Arabic, Burmese, Danish, Greek, Hebrew, Khmer, Lao, Swahili, Tagalog and the rest; nine Chinese regional varieties are listed separately, including Cantonese, Sichuanese and Hokkien. The licence is Apache 2.0, free for commercial use with no threshold — consistent with the rest of the current MiniCPM family. Installation is a single Python package, documentation is published on Read the Docs, and a hosted playground runs on Hugging Face Spaces. The repository records 453,089 downloads and 1,534 likes.

▸Videos

No videos yet.

▸Reviews