Sopro is a lightweight voice-cloning text-to-speech project from Halo NeuroAI, a startup that builds communication tools for people who lost their voice to ALS or post-stroke aphasia. The repository ships Sopro V2 Turbo, a 120-million-parameter model that streams audio, runs comfortably on a laptop CPU or entirely in the browser via ONNX, and clones a voice from just 5 to 20 seconds of reference audio. It supports English, French, German, and, notably, European Portuguese, a variant the developers say is, to their knowledge, not natively targeted by any other open TTS model, motivated by closed providers like OpenAI, ElevenLabs, and Cartesia mispronouncing it because of Brazilian-Portuguese data bias. The architecture pairs an autoregressive transformer predicting semantic tokens from an ASR-aligned tokenizer, warm-started from Whisper large-v3 with an FSQ bottleneck, with a flow-matching acoustic head generating mel spectrograms and a fine-tuned causal Vocos vocoder for streaming; the head runs in two solver steps via a distillation-plus-reflow pipeline, a 16x cut from the 32-step teacher with, per its own benchmarks, no measurable quality loss. On an Apple M3 CPU it reports an offline real-time factor of 0.24 and a streaming time-to-first-audio of about 300 ms; on a single H100 those drop to 0.07 RTF and roughly 200 ms. On Seed-TTS-eval test-en it reports a 1.63% word error rate for the 120M Turbo model, ahead of several larger systems it compares against including CosyVoice 2, VibeVoice-1.5B, and MaskGCT, though Qwen3-TTS scores lower at 1.7B parameters. It succeeds Sopro V1, a $250 hobby-budget model with unstable, inconsistent cloning; V2 was trained with donated compute from FCCN-FCT, first on Portugal's Deucalion supercomputer and later on MareNostrum 5 at the Barcelona Supercomputing Center. The maintainers deliberately skipped audio watermarking, arguing an open pipeline makes it trivial to strip and would only give a false sense of safety, and ask users not to impersonate people with it. Training code is not planned for release, though inference code, pretrained weights, a PyPI package, an npm ONNX package, and a CLI are all available today.