April 13, 2026

OpenBMB unveils VoxCPM 2, a 2B-parameter TTS model with 30 languages support

OpenBMB has released VoxCPM 2, a cutting-edge text-to-speech (TTS) model featuring 2 billion parameters and designed for multilingual voice synthesis across 30 languages. This model advances the previous VoxCPM releases by enhancing long-text stability and supporting zero-shot voice creation without a reference audio, making it suitable for professional applications such as filmmaking and gaming. VoxCPM 2 is capable of producing high-quality, expressive speech with features like controllable voice cloning and a unique voice design that utilizes natural-language descriptions to craft new voices. This evolution is part of a broader trend in TTS technology aiming to provide seamless and realistic multilingual capabilities.

OpenBMB: OpenBMB, or Open Lab for Big Model Base, is an AI research organization dedicated to developing open-source foundation models and systems advancing towards artificial general intelligence. The lab produces projects like ChatDev and multimodal models such as MiniCPM. In this news, OpenBMB launched VoxCPM 2, their latest tokenizer-free TTS model designed for production-grade multilingual speech synthesis with advanced voice cloning features.
VoxCPM 2: VoxCPM 2 is an open-source text-to-speech model employing a tokenizer-free diffusion autoregressive architecture built on a MiniCPM backbone for generating natural, expressive speech in continuous latent space. It enables zero-shot voice design from natural language descriptions, controllable voice cloning from short audio clips, and high-fidelity continuation cloning while producing studio-quality audio. The release introduces expanded multilingual capabilities including Chinese dialects and Southeast Asian languages, positioning it for professional applications in content creation and voice agents.
Apache-2.0: Apache-2.0 is a permissive open-source software license that allows free use, modification, and distribution for both commercial and non-commercial purposes with minimal restrictions. It requires preservation of copyright notices and disclaimers in distributions. In the news, VoxCPM 2 is released under the Apache-2.0 license, enabling broad commercial deployment.

Model Evolution: Evolves from initial VoxCPM releases in late 2025 with improvements in long-text stability, fine-tuning support, and now unified multilingual voice generation.
Community Integration: Rapidly adopted in tools like ComfyUI for node-based voice cloning and LoRA training workflows.
Benchmark Competitiveness: Delivers state-of-the-art results on zero-shot TTS evaluations like Seed-TTS-eval and multilingual tests against models from CosyVoice and Qwen.

Sources