Seed LiveInterpret 2.0

Seed LiveInterpret 2.0 (July 2025, arXiv 2507.17527) is ByteDance Seed's end-to-end simultaneous interpretation model — the rare SI system presented as a finished product rather than a lab demo. It uses a duplex speech-to-speech understanding-generating framework trained with large-scale pretraining plus reinforcement learning, and it attacks the exact problems that sink real interpreters: multi-speaker confusion, translated-speech inflation on long discourses, and the lag of cascaded pipelines. The headline numbers are all ByteDance's: over 70% correctness in complex scenarios as judged by human interpreters, quality beating commercial SI solutions by a wide margin, and cloned-voice latency cut from ~10 seconds to ~3 seconds. The catch is scope — the documented language pair is Chinese↔English, and the model itself isn't open; the public touchpoint is a demo on ByteDance's Volcano Engine Ark console. Still, it's the clearest signal of where industrial simultaneous S2ST stands in 2025.

• >70% correctness in complex scenarios, judged by human interpreters (author-reported)\n- Outperforms commercial SI solutions on translation quality (author-reported, RealSI benchmark)\n- Latency of cloned speech cut ~10s to ~3s (author-reported)\n- No independent benchmark replication found