FIGURE 10.

The SR resonance effect in different network architectures using the FSDD data set. (A) The plot shows the test accuracy as a function of the applied hearing loss for a deep convolutional network architecture (dark blue, starting at 400 Hz, exact network architecture shown in Table 2) trained on English words (digits: 0–9). The impaired speech comprehension by the hearing loss can be partly compensated by adding Gaussian noise (stochastic resonance). The cyan curve shows the improvement of speech comprehension for the optimal noise level (maxima values in panel B). (B) Test accuracy for different hearing losses (shades of blue) as a function of the added noise. The maxima show that SR can help to restore speech comprehension. (C) Similar analysis as shown in panel (A) for a two layer LSTM network (exact network architecture shown in Table 3); (D) Similar analysis as shown in panel (B) for the LSTM architecture. The improvement of speech perception in impaired systems (hearing loss) is a universal principle and does not depend on the used neural network.