Halo NeuroAI Releases Sopro V2 TTS Family, Open-Sources Turbo Model
Start here
Halo NeuroAI released Sopro V2, a family of text-to-speech models, and open-sourced its fastest member, sopro-v2-turbo, a 120M-parameter voice-cloning model that runs on a laptop CPU or in the browser and covers English, German, French and European Portuguese. The company reports that on an Apple M3 CPU it generates offline at 0.24 RTF and streams at 0.21 RTF with about 300 ms to first audio, while a single H100 reaches 0.07 RTF offline. The model was trained in four stages, including distillation from a 0.5B teacher and a reflow step that cut acoustic solver steps from 32 to 2; Halo NeuroAI reports 1.63 WER on Seed-TTS test-en and 1.79 on LibriSpeech test-clean under the F5-TTS protocol. A local demo starts with one command, and a fully in-browser ONNX demo needs no server, though on mobile the model is quantized and results may be slightly worse, and low-memory devices may crash. Halo NeuroAI says it does not plan to release the training code soon because of its complexity.
Generated from reports · updated 2 h ago
Report timeline
Follow the reports to see the event from different sides.
- Hacker News · AIHalo NeuroAI open-sources Sopro V2 Turbo, a 120M on-device voice-cloning TTS model
Halo NeuroAI released Sopro V2, a family of text-to-speech models, and open-sourced its fastest member sopro-v2-turbo, a 120M-parameter voice-cloning model that streams on a laptop CPU or in the browser and covers English, German, French and European Portuguese. On an Apple M3 CPU it generates offline at 0.24 RTF and streams with about 300 ms to first audio at 0.21 RTF, while a single H100 reaches 0.07 RTF offline. The model was built through four training stages, including distillation from a 0.5B teacher and reflow that cut acoustic solver steps from 32 to 2, and the authors report 1.63 WER on Seed-TTS test-en and 1.79 on LibriSpeech test-clean under the F5-TTS protocol.
Heat for this event
Not enough continuous observations yet to draw a trend.