Generalization to non-English code-switched language pairs

Determine how utterance-level code-switched language-identification systems can generalize from English-involved training pairs to unseen code-switched pairs involving only non-English languages.

Background

The CS-FLEURS XTTS2 subset consists entirely of unseen language pairs that do not contain English, whereas the training data is dominated by English-involved code-switching. The models perform almost perfectly on many seen synthetic pairs but obtain near-zero accuracy on XTTS2 unseen pairs.

This result indicates that the main unresolved issue is not merely representing unseen combinations of individually known languages, but transferring language-set prediction across the distributional and linguistic shift from English-involved training pairs to non-English pairs.

References

In contrast, the near-zero accuracy across all models on XTTS2 unseen pairs demonstrates that generalizing from English-involved training data to non-English CS pairs remains an open challenge.

— Code-Switching Spoken Language Identification as Multi-Label Set Prediction  (2610.01450 - Mitsumori et al., 1 Oct 2026) in Section 4, Results, paragraph “CS-FLEURS subset analysis”