Robust language-set prediction for diverse code-switched speech

Establish robust utterance-level code-switched language-set prediction for unseen language pairs, non-English code-switching, and real-world speech without relying on oracle language cardinality or unstable thresholds.

Background

The paper formulates utterance-level code-switched spoken language identification as multi-label language-set prediction, in which a system must identify all languages present in an utterance rather than assign an atomic pair label. This formulation is intended to represent language combinations not observed during training and to avoid requiring the number of languages in advance.

The experiments show that atomic-pair classifiers fail on unseen combinations, score-based systems perform best when given oracle cardinality, and the proposed set generator does not yet match the oracle Top-k baseline. The authors therefore identify robust prediction across unseen, non-English, and naturally occurring code-switching conditions as an unresolved challenge.

References

Robust language-set prediction for unseen, non-English, and real-world CS speech thus remains an open challenge.

— Code-Switching Spoken Language Identification as Multi-Label Set Prediction  (2610.01450 - Mitsumori et al., 1 Oct 2026) in Section 1, Introduction

Extending utterance-level CS-LID beyond two-language mixtures thus 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 “Three-language code-switching”