VoxCPM2: Text-to-Speech That Speaks 30 Languages Without Gimmicks

Here's the thing: speech synthesis has been obsessed with discrete tokenizers for years. Take text, convert to tokens, generate voice tokens, reconstruct audio. It's efficient, sure, but VoxCPM2 decided to flip the script.

The system works without tokenizers, generating speech representations directly via a diffusion-autoregressive end-to-end architecture. Sounds complicated? In practice, it means the model "thinks" directly in sound, skipping that intermediate token translation step. Result: more natural, expressive voices trained on over 2 million hours of multilingual data.

And those 30 supported languages aren't marketing speak—they actually work. Chinese to Turkish, Thai to Portuguese, plus Chinese dialects. Write in any language and the model synthesizes directly.

But the real magic is three features that change the game. First: Voice Design. Describe a voice with words—"a young woman with a gentle, sweet voice"—and the model creates it from scratch. No reference audio needed, just your imagination driving the output. Second: Controllable Cloning. Give a 5-10 second audio clip and clone the voice, with options to steer emotion, pace, and expression while keeping the original timbre. Third: "Ultimate Cloning" mode for perfectionists—provide both audio and transcript, and the model continues exactly where you left off, reproducing every vocal nuance.

Output is 48kHz studio-quality audio processed by AudioVAE V2 with asymmetric design—no external upsampler needed. And it's fast: Real Time Factor of 0.3 on RTX 4090, 0.13 with Nano-vLLM acceleration.

One more thing: it's all open-source under Apache-2.0, free for commercial use. Code and weights on GitHub, playground ready to test right now.