Sopro V2 Turbo: 120M-param voice-clone TTS, 5x real-time on CPU, Apache-2.0
SammyDaBeast · reddit · 2026-10-04
An indie developer shipped Sopro V2 Turbo 2610, an interim update to the lightweight open-source voice-cloning TTS: reduced roughness/break-up on previously problematic cloned voices, same 120M model at 300ms first-audio latency and 5x real-time on a laptop CPU, Apache-2.0 licensed, supporting English, European Portuguese, French, and German. Known weaknesses: very high-pitched/cartoonish voices and noisy reference audio. One-command local run via uvx; weights, browser demo, and HF Space all available.
Related event: Open-source TTS Sopro V2 Turbo 2610 fixes voice cloning artifacts(2 posts)→
More from Multimodal
- Topaz launches Starlight Fast 3: 4x faster video upscaling, 480p to 1080p in ~90 seconds — TinfoilTricorn · 2026-10-04
- Claude-generated animation channels explode on TikTok: 122k followers in 5 days — rickasaurus · 2026-10-04
- The LoopHole Project opens AI short film collab: 20-second clips chained into one community movie — LudovicCreator · 2026-10-04
- Community finds more DMAD steps nearly eliminate shimmering in Minimax H3 video generation — Devajyoti1231 · 2026-10-04
- Armaan Bansal unveils Future Archive 2026, an Indian superhero visual series — ivanhzhao · 2026-10-04
- Letting Claude prompt Midjourney: machines telling machines how to make pictures — technollama · 2026-10-04