GSLM2 multi-stream generative spoken model proposed

Researchers have introduced GSLM2, a multi-stream generative spoken language model designed to overcome the linguistic and data scaling bottlenecks of traditional single-stream speech models. Using an encoder called SpidR 2.0 and a shared temporal backbone, GSLM2 learns complementary speech properties at multiple levels of abstraction to model cross-stream dependencies. The authors report that the system consistently improves linguistic competence, narrows the gap with text-based language models by about two orders of magnitude across several tasks, and outperforms competitive speech-text systems. Open article →