●builderThis methodology offers a blueprint for generating high-quality, verified synthetic data for multilingual speech tasks.
●researcherThe use of hybrid synthesis and GSPO for audio-dependent reinforcement learning provides a new template for speech-language model training.