VoxCPM2: open 2B-parameter TTS model covering 30 languages

The open-source VoxCPM2 speech synthesis model has 2 billion parameters, was trained on over 2 million hours of audio and supports 30 languages including Thai. In the team's own MiniMax-MLS benchmark it scores a WER of 2.96 for Thai versus 73.94 for ElevenLabs, but trails badly in Arabic, Czech and Romanian.

• 2B parameters, 2M hours of audio, 30 languages including Thai
• Thai WER 2.96 vs 73.94 for ElevenLabs in team's tests
• Arabic WER 13.05 vs 1.67 for ElevenLabs and Minimax
• RTF about 0.3 on RTX 4090, down to 0.13 with vLLM-Omni