Sopro V2 Turbo 2610: cleaner voice cloning, 120M model, ~300ms first audio on CPU

SammyDaBeast · reddit · 2026-10-04

Open-source voice cloning model Sopro V2 Turbo ships the 2610 interim update, mainly fixing roughness and break-up on cloned voices.

Key points:

Same 120M model, same speed: 300ms to first audio on a laptop CPU, with true streaming
Apache-2.0; supports English, European Portuguese, French, German; more languages and voice controls planned
Known weaknesses: very high-pitched/cartoon voices, noisy reference audio, some OOD voices; author soliciting failed samples
One-line local run via uvx, plus repo, HuggingFace weights, in-browser demo, and an evals blog post

A fit for those who like F5-TTS but want a much lighter model that streams and runs comfortably on CPU.

More from Multimodal