Voice cloning models, measured across five languages

Short summary

Comparative benchmark of four open-source voice-cloning models (OmniVoice, Chatterbox, VoxCPM2, Fish Audio S2 Pro) across five languages using speaker similarity, error rates, and inference speed. OmniVoice delivered strongest overall performance; VoxCPM2 excelled on Arabic; Fish Audio S2 showed high similarity but slower inference. Published as engineering benchmark using Google FLEURS, not human evaluation.

• •OmniVoice int8 was strongest all-around model
• •VoxCPM2 bf16 excelled on Arabic speaker matching
• •Fish Audio S2 Pro fastest on German/Arabic but slowest inference speed

Generated with AI, which can make mistakes.