Code-Switching Detection via Language-Selective Dimensions in Speech Foundation Models

Apple researchers have developed a method to identify code-switching, where speakers alternate between languages, within large speech datasets. Because these patterns are currently underrepresented in training data, this technique allows developers to automatically extract and include relevant utterances for better model performance. This advancement addresses a significant gap in multilingual speech processing, as current foundation models often struggle to accurately transcribe or interpret spontaneous shifts between languages.