VoxCPM

VoxCPM2 is the tokenizer-free TTS model that makes traditional text-to-speech look like a toy. Most TTS systems rely on tokenizers—they break text into discrete units, then convert those units back to speech. That middleman step introduces errors, limits expressiveness, and honestly sounds robotic in ways users have learned to accept. VoxCPM skips all of that. It generates speech directly from raw text, which means multilingual support comes without the usual language-specific tokenizer headaches. Want to clone a voice? VoxCPM2 does true-to-life voice cloning. Want to design a completely creative voice that never existed? It handles that too.

OpenBMB built this. They're the team behind the CPM large language model series, and they specialize in making open-source AI tools that actually work. VoxCPM2 is their answer to a simple question: why are we still pretending that tokenization is necessary for good speech synthesis? The answer is: it's not. And once you hear VoxCPM2, you'll wonder why everyone's been doing it the hard way.

It's on GitHub, it's open, it's tradable. This is the kind of project that makes the closed-source TTS giants nervous.