开发者基于面壁智能VoxCPM2模型，构建了开源猫娘语音数据集NekoAudio。该数据集包含80K短音频（10-30秒，约100GB）和30K长音频（约200GB），去除对话文本中的舞台指示，专为TTS微调与角色语音建模设计。开发者已用NekoAudio-80K微调VoxCPM2-2B并开源演示模型NekoCPM2_demo。
🎙️ Build with VoxCPM2: NekoAudio — Let Your Catgirl AI Speak 🐱✨
What if your AI character could truly have a voice?
Developer @MindLiu67966 created NekoAudio, an open-source catgirl audio dataset designed for character voice modeling and TTS fine-tuning.
Built from NekoQA-30K, the dataset:
🐾 Removes stage directions/action descriptions from dialogue text for more natural speech
🎧 Includes 80K short audio samples (10–30s each) optimized for TTS fine-tuning
🗣️ Provides audio-text pairs for exploring character voice adaptation and expressive TTS
The dataset includes:
📦 Neko_Audio-80K_Short — ~100GB short audio dataset for efficient fine-tuning
🔗: https://huggingface.co/datasets/liumindmind/Neko_Audio-80K_Short
📦 Neko_Audio-30K_Long — ~200GB long-form audio dataset for extended speech generation
🔗: https://huggingface.co/datasets/liumindmind/Neko_Audio-30K_Long
The creator also fine-tuned VoxCPM2-2B with NekoAudio-80K and open-sourced the demo model NekoCPM2_demo.
A fun exploration of how open-source voice models can bring AI characters to life. 🚀
Powered by VoxCPM2:
🤗 Hugging Face: https://huggingface.openbmb.com/model/openbmb/VoxCPM2
来源：OpenBMB · x.com