The CREMA-D experiment—and why it stayed separate

CREMA-D contains acted emotional speech with repository-provided labels. It offered a tempting controlled question: can a fresh adapter over Base Inkling use audio to recover an emotion label or expressive description?

The appeal was methodological. Public metadata reduces ambiguity about the intended class, the clips are short, and a held-out actor can create an untouched evaluation split.

Seventy-seven repository-labeled clips rendered into a fresh training set. The project explicitly accepted the metadata at face value instead of pretending every recording had received a new human listen. The fresh rank-32 adapter trained for 20 optimizer steps and saved final checkpoints.

A separate comparison workflow prepared paired base-versus-adapter responses and a blind review interface. Training completion still did not establish better audio perception.

A model can improve on one and fail on the other. Combining the datasets without a carefully designed multi-task objective would make the result harder—not easier—to interpret.

The detour clarified three things. First, training must start from an explicit parent identity. Second, metadata-supervised and human-reviewed data are different evidence classes. Third, an evaluation dataset must match the behavior we want to ship.