OpenBMB/VoxCPM2 at 18.7K Stars: Tokenizer-Free TTS With 2B-Param Diffusion Autoregressive Model Generates 48kHz Audio in 30 Languages

GitHub·medium signal

VoxCPM2 skips discrete tokenization entirely, generating continuous speech representations through an end-to-end diffusion autoregressive architecture (LocEnc → TSLM → RALM → LocDiT). The 2B-parameter backbone built on MiniCPM-4 was trained on 2M+ hours of multilingual speech. Achieves 1.84% WER on Seed-TTS-eval English and 79.5% speaker similarity. Supports voice design from natural language descriptions.