Generalization of seed variability across speech architectures

Determine whether the training-seed-driven fairness variability observed in warm-start adaptation of Granite-Speech models also occurs in a second speech-LLM architecture outside the Granite-Speech family.

Background

The study evaluates seed variability primarily within the Granite-Speech architecture family, with an additional 8B replication that changes decoder scale but not the broader architecture family. The authors therefore leave unresolved whether the observed fairness instability is specific to this architecture family or represents a more general property of speech-LLM adaptation.

Resolving this question would test the external validity of the reported finding that training seeds can affect demographic fairness metrics more strongly than the studied compression variable.