OmniVoice vs VoxCPM2: which open text-to-speech model should you use?

OmniVoice covers 600+ languages but its weights are non-commercial; VoxCPM2 covers 30 languages under Apache 2.0. Compare languages, voice design and license.

  • Updated

Verdict. Choose by license first, then by languages. VoxCPM2 is a 2B model under Apache 2.0, so you can use it in a product. It covers 30 languages with 48 kHz output, voice design from a text description and cloning, and it runs in about 8 GB of VRAM. OmniVoice is smaller, at 0.6B, and covers 600+ languages, but its weights are CC BY-NC, so it is for research and personal projects even though the code is Apache 2.0. Voice design in OmniVoice was trained on Chinese and English only, and both developers forbid impersonation and fraud.

Specifications

Specifications OmniVoiceVoxCPM2
DeveloperXiaomiOpenBMB
ReleasedApril 1, 2026April 2026
Total parameters613M (OmniVoice)2.29B (VoxCPM2)
Active parametersAllAll
ArchitectureDiffusionDiffusion
Context window––
InputText, AudioText, Audio
License CC BY-NC (Non-commercial) Apache 2.0 (Permissive)
Memory for the smallest variant (4-bit)3.75 GB (OmniVoice)8 GB (VoxCPM2)

Performance claims come from the developer and have not been independently verified.

Pick OmniVoice if

  • You need a language outside VoxCPM2's 30, since OmniVoice lists 600+ languages.
  • Your use is research or personal, so the CC BY-NC license is fine.
  • You want fast batch generation on one GPU from a small 0.6B model.
  • You need low-resource languages that other models do not cover, and can accept less stable results.

Pick VoxCPM2 if

  • You are building a product, because Apache 2.0 allows commercial use.
  • You want a new voice from a text description, with no reference audio.
  • You need 48 kHz output or Chinese dialects.
  • You want to clone a voice or fine-tune from a few minutes of audio under Apache 2.0.

OmniVoice · VoxCPM2