Voice Activity Detector for Device with Small Processor and Memory

Automatic Speech Recognition needs a lot of computation resources to do all operations necessary to recognize speech. Thus it could not be deployed on devices with small processor and memory. To overcome this problem, a new and simple Voice Activity Detector is needed. This is achieved by combining Raw Speech Signal as input and Deep Neural Network as classifier. Training data is prepared by collecting recorded speech from various speakers and assigning label on them manually. Minimum neural network architecture, in term of number of layer and number of neuron that still provides good detection accuracy, is determined using experiments. The result is an architecture with only 3 layers, 130 neurons, 64 inputs, softmax activation function, Adam optimization, dropout rate 0.2, and batch size 64. The accuracy is 0.7406 for training and 0.7168 for validation.