Halo Neuro, the Portuguese startup building voice-cloning tools for people who lost their speech to ALS or stroke-related aphasia, released a new text-to-speech model family called Sopro V2 and open-sourced its smallest member, a 120-million-parameter model called Sopro V2 Turbo. The model clones a voice from a short reference clip, streams audio in real time, and runs locally on a laptop’s CPU or directly inside a browser tab, no server required. That last detail is the story: voice cloning stops being something rented from an API and becomes something anyone can download and run on their own machine.
Sopro V2 Turbo generates speech on an Apple M3 CPU at a real-time factor of 0.24 offline, with streaming output starting in roughly 300 milliseconds, according to Halo Neuro’s own benchmarks. On a single H100 GPU, the same model reaches 200 milliseconds to first audio. Halo Neuro says the model is multilingual, covering English, German, French, and, notably, European Portuguese, which it describes as the first open TTS model built natively for that dialect rather than the Brazilian Portuguese most training corpora default to.
The open weights belong to Turbo alone. Sopro V2 itself is a larger 0.5-billion-parameter model that Halo Neuro trained first and has kept closed. Turbo is a 120-million-parameter student distilled from that teacher, then compressed further through a technique the company calls reflow, which cut the acoustic model’s solver steps from 32 down to two: a 16x speedup the company reports without any drop in quality or intelligibility it was able to measure.
The backstory explains the urgency behind the project. Halo Neuro’s paid product depends on commercial voice-cloning providers, among them OpenAI, Cartesia, and ElevenLabs, and the founders say none of them handled European Portuguese well, a gap they trace to Brazilian-heavy training data. An earlier version, Sopro V1, started as a two-week side project on a $250 training budget. It climbed to the number two spot on Hacker News but was unstable and English-only. Sopro V2 replaced most of that architecture: a convolutional core gave way to a transformer decoder, and the acoustic head moved from discrete codebooks to continuous mel-spectrogram prediction, the same approach used by F5-TTS and CosyVoice.
On Halo Neuro’s own benchmark tables, Turbo posts a 1.65 word-error rate on the Seed-TTS English test set, ahead of several closed and open models three to fourteen times its parameter count. That comparison rests entirely on Halo Neuro’s own evaluation runs rather than an independent leaderboard, and the company does not cite third-party verification of the numbers. Halo Neuro also discloses, in the same post, that it added no watermarking to the model’s output, arguing that an open-source inference pipeline would make any watermark trivial to strip out anyway.
That disclosure is worth sitting with longer than the benchmark table. A 120-million-parameter model that clones a voice from a few seconds of audio and runs on ordinary consumer hardware removes the two constraints that used to slow down impersonation: cost and infrastructure. Nobody needs a paid API account that logs requests or can refuse a suspicious one. Halo Neuro’s release asks users not to impersonate people, but a request in a README carries no technical weight.
For teams building voice products, the practical read is that on-device cloning quality has closed most of the gap with hosted providers using a fraction of the parameters those providers likely run. For anyone who treats a familiar voice as proof of identity, from bank call centers to family phone scams, the next ninety days are the window to stop accepting “it sounded like them” as evidence on its own.
Details and benchmark figures are from Halo Neuro’s research post “Sopro V2: private, fast, on-device text-to-speech,” published on research.haloneuro.ai.
