Open-source LuxTTS claims 150x real-time voice cloning in 1GB VRAM
thisguyknowsai · x · 2026-07-28
- An open-source TTS model called LuxTTS claims 150x real-time generation while fitting in 1GB VRAM.
- It distills ZipVoice down to 4 sampling steps and uses a custom 48kHz vocoder, aiming to avoid the thin, compressed sound common in many 24kHz TTS systems.
- The project says it runs faster than real time on CPU, supports Mac MPS, and is released under Apache 2.0.
- The post frames it as a voice-cloning model that keeps quality close to much larger systems while remaining lightweight enough for local use.
More from Multimodal
- Open-source local AI tool generates slides, edits them visually, and exports PPTX — goodboydhrn · 2026-07-28
- Mage-Flow runs 13–18× faster than Krea 2 Turbo on an RTX 3060, but quality trails — SirMick · 2026-07-28
- MIT framework teaches vision-language models to generate more accurate CAD programs — bravo_abad · 2026-07-28
- Text like “33°C” is not touch, argues a critique of LLM embodiment claims — flowersslop · 2026-07-28
- A ready-made Seedance 2.0 prompt recreates a 1990s arcade scene — techhalla · 2026-07-28
- Developer builds a local photo assistant with OpenClaw and MiniCPM-V 4.6 — 面壁智能 · 2026-07-28