| title: Offex Voice Studio CPU | |
| emoji: 🖥️ | |
| colorFrom: indigo | |
| colorTo: pink | |
| sdk: gradio | |
| sdk_version: 5.45.0 | |
| app_file: app.py | |
| pinned: false | |
| # Offex Voice Studio (CPU edition) | |
| The same voice-cloning studio as [Offexb/Ai](https://huggingface.co/spaces/Offexb/Ai), running fully on CPU. | |
| ## Why CPU? | |
| - **No daily GPU quota at all** — generate as much as you like. | |
| - The trade-off: each generation takes several minutes instead of seconds. | |
| ## How it works | |
| - IndexTTS-2 with the model snapshot and all auxiliary models downloaded by a background warm-up thread at startup. | |
| - Reference clips are auto-trimmed of silence, capped to 20 seconds and peak-normalised for a closer clone. | |
| - Two permanent reference voices ship with the Space (`assets/`), preloaded by default with quick-pick buttons. | |
| - Optional emotion reference, text-derived emotion and seed control. | |
| ## Runtime notes | |
| - The SDK is pinned to Gradio 5.45.0 (launched with `ssr_mode=False` for correct mobile rendering). | |
| - tensorboard is pinned to 2.18.0 for NumPy 2.x compatibility; app.py re-adds the removed `np.bool8` alias defensively for other legacy dependencies. | |
| - Keep texts short — CPU synthesis is slow. | |