Sopro V2 Turbo in-browser TTS · voice cloning

Acoustic flow model + vocoder run on WebGPU (fp16); the small autoregressive semantic LM runs on multithreaded WASM as in the official runtime, which benchmarks ~3x faster than per-token GPU readback. Flow steps are fixed at 2 in this Turbo export; the model has no speed control. Language is auto-detected unless forced. Reference crop changes re-encode the voice.