Lure-TTS and Lure-Turbo: Fast Open-Source 600M Voice Models
A developer has introduced Lure-TTS and Lure-Turbo, open-source 600M parameter voice models trained for under $200 that achieve 4x real-time inference speeds. The models support voice cloning, inline tagging for emotions or events, and directional voice control. This project demonstrates that high-performance, customizable voice assistants can be built efficiently on modest hardware without requiring massive industrial budgets. It significantly lowers the barrier for developers looking to run local, responsive text-to-speech models. Trained by fine-tuning a base model using techniques like multi-teacher on-policy distillation, Lure-Turbo optimizes a 15-pass code predictor down to 3 passes to achieve a real-time factor (RTF) of 0.14. The accompanying inference engine supports CUDA graphs, Mac (mlx-audio), and CPU (audio.cpp) backends for fast initial audio generation.
## BACKGROUND
On-policy distillation is a machine learning technique where a student model generates trajectories that are evaluated and guided by a teacher model to improve performance. Text-to-speech (TTS) optimization often focuses on reducing the real-time factor, allowing models to synthesize spoken audio faster than it takes to listen to it.