VoxCPM2
Description
VoxCPM2 is an open text-to-speech model from OpenBMB with 2 billion parameters. Per the model card, it speaks 30 languages without a language tag, outputs 48 kHz audio, designs a new voice from a text description (gender, age, tone, pace) and clones a voice from a short reference clip; with the clip's transcript it can continue the voice more precisely. It supports streaming output and fine-tuning (full or LoRA, from 5–10 minutes of audio). The authors say it was trained on more than 2 million hours of speech; we did not verify this. The card states about 8 GB of VRAM; speed is about 0.3 real-time factor on an RTX 4090 and about 0.13 with Nano-vLLM, according to the authors. The weights are public on Hugging Face without any acceptance step, released in April 2026 under the Apache-2.0 license, which allows commercial use. The authors offer an official demo Space on Hugging Face and a demo page with audio examples; code and docs are on GitHub. Voice cloning should only be used with the speaker's consent. Date of check: 2026-10-19.
Pros and cons
Pros
- 30 languages, 48 kHz output, per the model card
- Voice from a text description or from a short clip
- Streaming output and fine-tuning (full or LoRA)
- Apache-2.0 license, weights open without an acceptance step
- About 8 GB of VRAM; official Space for trying it
Cons
- Speed and training-data figures are the authors' own statements
- Not on the public speech arena we use, so quality is our expert estimate
- Cloning a real person's voice needs that person's consent
- Fewer languages than the largest multilingual models
Key facts
- License
- Apache-2.0
- Commercial use
- Allowed
- Hardware requirements
- About 8 GB of VRAM, per the model card (bfloat16)
- Developer
- OpenBMB
- Interface languages
- English
- Checked on
- Oct 19, 2026