LuxTTS: small-and-fast open voice cloning — 150x realtime on one GPU, native 48 kHz, 1 GB VRAM, fully local

A fresh open-source voice-cloning pick from blogger bkdgiffug (lumxss): LuxTTS, a lightweight zipvoice-based TTS model that goes "small and fast" — (1) speed: up to ~150x realtime on a single GPU, faster than realtime even on CPU (CUDA/MPS/CPU all supported); (2) clarity: native 48 kHz high-fidelity output where most TTS models stop at 24 kHz; (3) efficiency: runs in about 1 GB of VRAM so older GPUs work too. The small model clones voices on par with ones 10x larger (SOTA-level cloning). The key selling point: fully local inference — your voice never leaves your machine; pip-install and go, with HuggingFace Space/Colab for quick trials. Great for dubbing and agent voice output in local-first setups. Note: the tweet carries an image only (no video); content cross-verified via the fxtwitter API and the GitHub README.