VoxCPM2 — Tokenizer-Free Multilingual TTS

A tokenizer-free TTS from OpenBMB based on a diffusion autoregressive architecture. 2B parameters, trained on 2M+ hours, 30 languages, 48kHz output, built on MiniCPM-4.

Architecture: LocEnc → TSLM → RALM → LocDiT, AudioVAE V2 asymmetric 16kHz→48kHz.

Installation

pip install voxcpm  # Python ≥3.10, PyTorch ≥2.5, CUDA ≥12
model = VoxCPM.from_pretrained("openbmb/VoxCPM2", device="auto")  # cuda→mps→cpu
# torch.compile issues: optimize=False; HF mirror: export HF_ENDPOINT=https://hf-mirror.com

Models