OpenBMB releases VoxCPM2, a 2B Apache 2.0 TTS with voice design and 48 kHz output
VoxCPM2 is a 2B tokenizer-free TTS covering 30 languages, with voice design, cloning and 48 kHz output, under Apache 2.0.
OpenBMB released VoxCPM2, a 2B-parameter text-to-speech model built on the MiniCPM-4 backbone with a tokenizer-free diffusion autoregressive design. The card lists 30 languages that need no language tag, plus nine Chinese dialects. It accepts 16 kHz reference audio and outputs 48 kHz.
Three voice modes are described. Voice Design builds a voice from a text description alone. Controllable Cloning copies a voice from a short clip and can add style guidance for emotion and pace. Ultimate Cloning uses reference audio plus its transcript. The card quotes a real-time factor of about 0.30 on an RTX 4090, or about 0.13 with Nano-vLLM, and says it needs roughly 8 GB of VRAM. LoRA and full fine-tuning are supported with 5 to 10 minutes of audio.
The licence is Apache-2.0, which permits commercial use. The card says results are state of the art or competitive but gives no numbers, pointing to GitHub. It warns that Voice Design output varies between runs, quality varies by language, and very long or expressive inputs can be unstable.
The card gives no release date; the date used is the Hugging Face repo creation date.
A permissive licence and honest caveats. The tables live elsewhere.
Sources
Editor's first pass (AI-assisted)
- Real 4.5: Apache 2.0 weights, a stated 8 GB VRAM need and streaming are all on the card.
- Useful 4: Design, cloning and 48 kHz output across 30 languages suit real production audio.
- Notable 4: Over 1,700 likes on Hugging Face; permissive licence helps.
- Novel 3.5: Tokenizer-free diffusion autoregression on MiniCPM-4 is a different route from codec TTS.
Comments