OpenBMB@OpenBMB · X面壁把语音生成压进2B参数，支持30种语言加方言克隆，还附完整技术报告和Apache 2.0开源，做语音产品的可以直接拉代码跑起来了。
面壁智能 OpenBMB 发布 VoxCPM2 技术报告。该模型为最新语音生成模型，拥有 2B 参数，基于超 200 万小时多语言语音数据训练，支持 30 种语言和 9 种中文方言。具备自然语言语音设计、可控及高保真延续性语音克隆能力。技术报告涵盖架构设计、统一序列公式、AudioVAE 高保真语音重建、大规模训练评估，以及零样本和指令跟随 TTS 基准结果。采用 16kHz 语义编码 + 48kHz 波形重建，在公开 TTS 基准上达到 SOTA 或极具竞争力。模型权重、微调代码和推理工具以 Apache 2.0 开源。
🚀 VoxCPM2 Technical Report is now available on arXiv！
VoxCPM2 is the latest speech generation model in the VoxCPM family. Built with 2B parameters and trained on over 2 million hours of multilingual speech data， it supports 30 languages and 9 Chinese dialects， along with natural-language voice design， controllable voice cloning， and high-fidelity continuation-based voice cloning.
In this technical report， we provide a comprehensive overview of： 🔹 The VoxCPM2 architecture 🔹 A unified sequence formulation for speech generation and control 🔹 The design of AudioVAE for high-fidelity speech reconstruction 🔹 Large-scale multilingual training and evaluation 🔹 Benchmark results across zero-shot and instruction-following TTS tasks
With 16kHz semantic encoding and 48kHz waveform reconstruction， VoxCPM2 delivers high-quality speech generation and achieves SOTA or highly competitive performance on public TTS benchmarks.
To support open research and development， we have open-sourced the model weights， fine-tuning code， and inference toolkit under the Apache 2.0 license.
📄 Paper： https：//arxiv.org/abs/2606.06928 💻 GitHub： https：//github.com/OpenBMB/VoxCPM We hope VoxCPM2 helps advance the open-source multilingual speech ecosystem. Feedback， experiments， and contributions are always welcome！ 🔥 #AI #OpenSource #TTS #SpeechAI #VoiceAI #GenerativeAI #MachineLearning