Fig. 2

Teacher DNN outputs for inputs from the same class. Posterior probability outputs by a teacher DNN for two inputs belonging to the same tied CD state are shown with state IDs for the five states having the highest posterior probability shown next to their corresponding bars. Even though the center frame belongs to the state, the output distributions are markedly different for the two inputs from different speakers, revealing different notions of similarity to other states learned by the teacher. This information can be conveyed to the student through soft targets but not hard targets