VoxCPM is an open-source text-to-speech model family with 39,000+ stars on GitHub that generates highly natural, expressive speech — and can design or clone a voice from a short prompt or clip.
Developed by OpenBMB and built on the MiniCPM-4 backbone, VoxCPM takes a different technical route from most TTS systems: it is tokenizer-free, generating continuous speech representations directly with a diffusion autoregressive architecture instead of predicting discrete audio tokens. The current release, VoxCPM2, is a 2-billion-parameter model trained on over 2 million hours of multilingual speech, supporting 30 languages with 48kHz studio-quality output. Beyond plain synthesis it offers three creative modes: Voice Design creates a brand-new voice from a natural-language description of gender, age, tone and emotion; Controllable Cloning copies a reference voice while letting you steer its delivery; and Ultimate Cloning continues seamlessly from reference audio plus its transcript, preserving fine vocal detail. Streaming synthesis runs faster than real time on a single RTX 4090. The project is actively developed, with VoxCPM2 released in 2026.
VoxCPM is free and open source under the Apache License 2.0, including the VoxCPM2 model weights. You self-host it, so the only cost is GPU compute.
Developers building voice interfaces, dubbing and audiobook pipelines, and creators who need custom or cloned voices without paying per-character fees to a TTS API.