Fig. 6
From: Strategies for distant speech recognitionin reverberant environments
Comparison of different adaptation strategies. The figure plots the WER as a function of the number of epochs for different adaptation strategies, i.e., (1) retraining the weights and bias parameters associated with the last hidden layer of the DNN (‘last’), (2) retraining only the first layer (‘1st’), (3) retraining the first two layers (‘1st + 2nd’), (4) retraining all parameters (‘all’), and (5) training a global feature transformation (‘LIN’). The results are given for the RealData set of the development set, for (a) distant speech and for (b) speech processed with 1ch WPE