Sopro V2: SOTA-level voice cloning TTS at 120M params, open-source and CPU-friendly

Run it locally with one command (downloads the model, opens a small web UI): uvx --from sopro soprotts serve A year ago I posted Sopro V1 on this subreddit, a small TTS I trained for ~$250 as a side project. It was unstable, English-only, and had a lot of rough edges. Sopro V2 is a new version with more compute and research behind it. What it is: sopro-v2-turbo : 120M params, Apache-2.0, zero-shot voice cloning from 5-20 s of reference audio English, European Portuguese, French, German . To my knowledge the first open TTS natively targeting European (not Brazilian) Portuguese Streaming : ~300 ms time-to-first-audio on an M3 CPU at 0.21 RTF, and 0.07 RTF on an H100 On Seed-TTS-eval test-en it reaches 1.51-1.65 WER, competitive with models 3-14x its size (F5-TTS 1.83, CosyVoice 3 2.02, Spark-TTS 1.98). Full tables, audio samples, and the architecture/training story are in the blog post. Links: Blog post (evals, samples, how it works): research.haloneuro.ai/posts/sopro-v2 GitHub: github.com/samuel-vitorino/sopro Weights: huggingface.co/samuel-vitorino/sopro-v2-turbo Python: pip install sopro In-browser demo (desktop): samuel-vitorino.github.io/sopro