VoxCPM

VoxCPM comes out of OpenBMB, the lab backed by Tsinghua University's NLP group and ModelBest, the same organisation behind the MiniCPM line. VoxCPM2 arrived in April 2026 at 2B parameters, following VoxCPM1.5 in December 2025 and the original 0.5B release in September 2025. Weights and code are Apache 2.0 with no field-of-use rider and no user cap, which at this quality tier is uncommon enough to be the headline rather than a footnote.

The architectural claim is tokenizer-free synthesis. Rather than quantizing audio into discrete codes and predicting those, VoxCPM2 uses a diffusion autoregressive design that produces continuous speech representations directly, which is what lets it emit 48kHz where most open TTS models settle for 24kHz. Training used more than 2 million hours across 30 languages, with nine Chinese dialects covered explicitly: Sichuan, Cantonese, Wu, Northeast, Henan, Shaanxi, Shandong, Tianjin and Southern Min. On Seed-TTS-eval English it reports 1.84% WER and 75.3 speaker similarity. Real-time factor sits near 0.30 on an RTX 4090 and falls to roughly 0.13 through the Nano-vLLM path, on around 8GB of VRAM.

Two capabilities separate it from the reference-clip cloning everything else does. Voice design generates a speaker from a written description with no audio input at all, which removes the consent question that clip cloning raises whenever the reference belongs to a real person. Controllable cloning then holds a cloned timbre while steering delivery through style instructions. The honest caveats are worth stating plainly: output varies between runs and the model card suggests generating several takes, which complicates any automated pipeline; the 30-language figure spans a wide quality range with Mandarin and English clearly best served; and the published install path assumes CUDA 12 with PyTorch 2.5 or later, so neither Ascend nor CPU is a first-party target. Weights are on Hugging Face and ModelScope, and both pull cleanly.