CASA Achieves SOTA Speaking Assessment with Whisper and Qwen
August 14, 2026
CASA combines Whisper-medium with Qwen3.5-2B to achieve a 0.358 RMSE on the Speak & Improve Corpus 2025. The architecture separates acoustic and content information while using approximately half the inference parameters of previous methods.
HOW THIS AFFECTS YOU
●
builderYou can deploy a more efficient, interpretable multimodal assessment tool.
●
researcherThe separation of speech delivery and content provides a clearer analysis of acoustic vs. semantic contributions.