OpenBMB/VoxCPM2: Tokenizer-Free 2B-Parameter TTS with 30-Language Voice Design — 7.4K Stars, +460 Today

GitHub·medium signal

VoxCPM2 introduces a tokenizer-free, diffusion autoregressive architecture (LocEnc → TSLM → RALM → LocDiT) that operates directly in continuous latent space, eliminating information loss from audio quantization. The 2B-param model supports 30 languages, voice design from natural language descriptions (no reference audio needed), and real-time streaming at RTF ~0.3 on RTX 4090. Trained on 2M+ hours of multilingual speech with 48kHz studio-quality output, it achieves 1.84% WER / 75.3% similarity on Seed-TTS-eval English benchmark. Apache 2.0 licensed, ~8GB VRAM.