Fig. 5From: A depthwise separable convolutional neural network for keyword spotting on an embedded systemSingle-inference test accuracy of networks trained on clean and noisy speech. Networks were tested in both noise set 1 (matched) and noise set 2 (mismatched). The number in parentheses specifies the average accuracy in the range from 0 to 20 dB SNR. The performance of the baseline model trained on noisy data is shown for comparisonBack to article page