LuxTTS is a text-to-speech model designed for high-quality voice cloning that achieves synthesis speeds exceeding 150x realtime.
The tool addresses the challenge of performing realistic voice cloning efficiently by using a lightweight architecture based on ZipVoice principles but distilled to four inference steps with an improved sampling technique. It generates clear 48kHz speech output, a higher quality than most TTS models limited to 24kHz, while maintaining a memory footprint under 1GB of VRAM. The approach enables the model to run on consumer GPUs and even faster than realtime on CPUs.
Developers should choose this tool if they need voice cloning capabilities with minimal computational overhead and want to run inference locally without cloud dependencies. It suits projects requiring high-quality speech synthesis at scale or on resource-constrained hardware. The README distinguishes it from ZipVoice by highlighting the distillation to fewer steps, improved sampling, and the custom 48kHz vocoder. The tool requires a minimum 3-second audio file for voice cloning and offers tuning parameters like return_smooth for addressing metallic artifacts and t_shift for balancing pronunciation accuracy against output quality.
The project maintains an active ecosystem with community-contributed implementations including Gradio interfaces, ComfyUI nodes, and ONNX variants. Development roadmap items indicate planned releases for version 1.5 and float16 inference optimization, suggesting ongoing work to improve performance further. The codebase is available with pre-trained models hosted on Hugging Face alongside interactive demo spaces and Colab notebooks for immediate experimentation.
