Explain the anomalous behavior of four target languages under language-subspace steering

Investigate why steering the CosyVoice3 text-to-speech model's language subspace toward Korean, Zulu, Kinyarwanda, or Vietnamese causes the synthesized output to degenerate, including substantially increased duration and phone and phone-feature error rates, when the behavior is not clearly explained by training-data volume or pretraining language coverage.

Background

The paper evaluates causal steering of a 15-dimensional language subspace in three phoneme-level speech models, including the CosyVoice3 text-to-speech model. Steering generally reproduces phonological interference by removing phonemes associated with the source language and introducing phonemes associated with the target language. However, steering CosyVoice3 toward Korean, Zulu, Kinyarwanda, or Vietnamese produces unusually degraded speech: the edited audio becomes substantially longer, and both phone error rate and phone-feature error rate increase relative to other target languages. Because these four targets exhibit a distinct failure mode that is not accounted for by the examined explanations involving training-data volume or pretraining language coverage, the authors leave its cause unresolved.