Lure-TTS: 600M voice models trained for under $200, running 4x faster than real time
smirkishere · reddit · 2026-10-25
An indie developer finetuned Qwen3-TTS 0.6B into Lure-TTS and a faster Lure-Turbo, for under $200 total on GPUs no bigger than an RTX PRO 6000.
Training pipeline:
- Trained classifiers to label intensity, events (laughs, cries, sighs), emotion, and direction, used to clean YODAS and Open Yap data;
- SFT stage taught the model to emit inline tags instead of reading them aloud;
- Multi-teacher OPD: separate teachers per skill (voice cloning, direction-following, tagging), merged via on-policy distillation;
- Turbo compresses the 15-pass code predictor to 3 passes, achieving RTF 0.14 vs 0.195 for the 5B Fish S2 Pro while matching speaker similarity (ECAPA 0.699) — 4x faster than real time.
Open-source inference engine: built on faster-qwen3-tts (CUDA), mlx-audio (Mac), and audio.cpp (CPU). On a 4090 it cuts first-audio latency from 10.6s to 0.17s and runs 5x faster overall (RTF 0.53 vs 2.54) via 0.32s streaming chunks, CUDA graphs, and batching up to 8 requests. One Python API/CLI/OpenAI-compatible server runs Lure, Turbo, and every Qwen3-TTS 12Hz checkpoint on Mac, NVIDIA, or CPU.
Training code and engine are on GitHub; the author is improving long-line tags, emotion switches, and freeform paraphrases.
More from Multimodal
- One prompt, four approvals: Claude produces a 78-second cartoon explainer video — alex_verem · 2026-10-25
- Canva's new Craft image model plans layouts first and splits images into editable layers — alex_verem · 2026-10-25
- Local RTX 5090 test: FLUX.1 Dev keyframes + Wan 2.1 14B I2V nail neck folds and wet-scale physics — LatentForge-Dev · 2026-10-25
- Credit: the Qwen3.5-4B 3D generation model is by community dev ccore2 — victormustar · 2026-10-25
- MCP turns AI podcast production into a workflow — the hard part is listener retention — aziz4ai · 2026-10-25
- Fully AI-made Arabic podcast built on GPT ASTRA 6 with MCP chaining Runway and ElevenLabs — aziz4ai · 2026-10-25