SyRA-CS Meta-Learning Framework for Code-Switched ASR Adaptation
October 23, 2026
SyRA-CS uses Reptile meta-learning to prepare pretrained models for code-switched speech adaptation using synthetic data. The framework improves Word Error Rate (WER) across Arabic, French, German, and Hindi-English pairs when adapting to just one hour of labeled real target speech.
HOW THIS AFFECTS YOU
●
builderThis offers a way to improve ASR performance in multilingual environments with limited real-world data.
●
researcherYou can leverage meta-learning to bridge the gap between synthetic and real-world speech data.