VoxCPM

0.5B TTS model by OpenBMB on a MiniCPM-4 backbone — about 5 GB of VRAM, RTF 0.17 on an RTX 4090, Apache-2.0. Twenty-three cards here carry a recipe from an 8 GB floor. An RTF well under 1.0 means it synthesises faster than real time by roughly six times, which is what makes it usable in a loop rather than as a batch step.