New Insights into Optimal Alignment of Acoustic and Linguistic Representations for Knowledge Transfer in ASR