VoxCPM2: Tsinghua's 2B open TTS with 30-language cloning at 48kHz studio quality

The VoxCPM series from Tsinghua’s NLP lab (OpenBMB) topped GitHub and HuggingFace Trending twice with earlier releases, and VoxCPM2 (April 2026) is the flagship: 2B parameters trained on over 2 million hours of multilingual speech, 30 languages, newly added Voice Design and controllable cloning, and 48kHz studio-quality output (Apache-2.0, 41,134 stars as of November 7; arXiv technical report 2606.06928). Deployment note: official guidance calls for at least 24GB of VRAM to run smoothly — a plain office PC will not carry it, so budget for an inference card or cloud GPU.

Core features

• Tokenizer-free architecture: continuous speech modeled directly, on a MiniCPM-4 backbone.
• 30 languages + voice design + controllable cloning: beyond standard TTS, describe the voice you want to generate a timbre, or clone from reference audio under controllable settings.
• 48kHz studio quality: output specification aligned with professional audio workflows, so post-production starts without an upsampling penalty.
• Real-time streaming: RTF as low as ~0.3 on an RTX 4090, ~0.13 accelerated by Nano-vLLM or vLLM-Omni — official omni-modal serving with PagedAttention and an OpenAI-compatible API.
• Fine-tuning friendly: SFT and LoRA scripts open-sourced since VoxCPM1.5.

Typical use cases

• Audiobook and podcast production: batch localization of multilingual spoken content.
• Voice asset design: brand or character voices designed and cloned under control.
• Real-time voice chat: streaming output plugs into realtime pipelines — RTF < 1 sustains it, and the 4090’s 0.3 leaves 3x headroom for batching.

Quick start

Python 3.10+, pip install and pull weights from HuggingFace (openbmb/VoxCPM2); an RTX 4090-class card runs streaming inference, lower-spec setups go cloud. A HuggingFace Spaces Playground lets you try before installing. The lineage is worth knowing: the 0.5B release (September 2025) and 1.5 (December 2025) each topped GitHub/HuggingFace Trending, and VoxCPM2 doubles down with 30 languages, voice design and cloning, and the 48kHz upgrade — weights, docs site, playground and an arXiv technical report all in place. Compared with the older 0.5B, the jump is not just languages: tokenizer-free modeling on the MiniCPM-4 backbone is what makes voice design (describing a voice in words) practical at all.

Summary

VoxCPM2 fits teams with GPU resources that need locally deployed, high-quality multilingual TTS; the 24GB VRAM floor keeps ordinary office PCs out — for lower budgets, community variants exist (including a 4GB Turkish-language wrapper). Facts and limits: Apache-2.0, Python, 2B parameters; voice-cloning responsibility stays with the caller — cloning someone’s voice requires their consent. VoxWeft on this site takes the on-device low-latency route for interpreting, VoxCPM2 the server-grade fidelity route — pick by deployment conditions.