VoxCPM Review 2026

Tokenizer-free multilingual TTS from OpenBMB — true-to-life voice cloning and creative voice design from a small open model.

VoxCPM is the open-source text-to-speech model line from OpenBMB (the team behind MiniCPM) that has become one of the most credible open alternatives to ElevenLabs and OpenAI's voices in 2026. The latest VoxCPM2 release is tokenizer-free — it directly models the speech signal with a continuous representation rather than going through a discrete acoustic codebook — which gives it noticeably more natural prosody, more accurate cross-lingual transfer and very high-quality voice cloning from short references. The model handles 'creative voice design' (describe a voice in text and have it synthesised), multilingual generation across Chinese, English and several other languages, expressive emotional control, and on-device deployment on relatively modest hardware. Because it is released under a permissive licence and ships pre-trained checkpoints, VoxCPM has been quickly integrated into open voice-agent stacks like Dograh and AIRI, used as the TTS layer for self-hosted assistants, and embedded into local apps for accessibility and game audio. It's a strong example of the small-but-capable open TTS wave that has emerged alongside larger closed models and has made high-quality voice cloning available to anyone with a workstation GPU.