Scaling CookVoice to larger models and datasets

Determine the scaling potential of CookVoice by evaluating larger Diffusion Transformer variants and substantially larger human-voice training datasets than the current 43.51-million-parameter model trained on 168 hours of data.

Background

CookVoice is currently implemented with the DiT-S diffusion-transformer variant and trained on a comparatively small 168-hour dataset. Although the framework achieves broad task coverage and efficient inference, the paper states that it has not yet been scaled because of resource constraints. Establishing how performance, controllability, and efficiency change with larger model capacity and training data remains unresolved.