Fig. 3From: Wise teachers train better DNN acoustic modelsProbability function for assigning speech/non-speech frames. Probability function that assigns probability p=f(x) that frame contains speech for each value of acoustic data x. When a training set is sampled from this distribution, similar values of x are likely to be mapped to different 0/1 (non-speech/speech) targets in the shaded region. The resulting function will be highly nonlinear in this region and difficult to learn using hard targetsBack to article page