VoxCPM2 is OpenBMB's 2B-parameter speech generation model, built on MiniCPM-4 and operating in the continuous latent space of AudioVAE V2. A local encoder, text-semantic language model, residual acoustic language model and local diffusion transformer generate latent patches that decode natively to 48 kHz audio. The project supports multilingual synthesis, natural-language voice design, reference-audio cloning, continuation cloning with transcript and LoRA or full fine-tuning. This deep dive separates implementation facts from author-reported benchmarks and examines streaming deployment, voice consistency and misuse risk.
As an Amazon Associate, we earn from qualifying purchases.
