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

This post proposes a lightweight method for code-switching detection directly from speech by leveraging language-selective dimensions in pretrained speech encoders. It demonstrates that pretrained speech encoders already encode the information needed to identify CS utterances and provides a tool for building more representative multilingual corpora. The Whisper-small encoder achieves a mean F1 of 0.798 across 14 language pairs in CS-FLEURS.