Abstract:
The Volterra series is a mathematical tool widely used to analyze and model nonlinear systems. The Volterra model expands a nonlinear system's response in terms of a series of integral equations. Like linear convolution, nonlinear convolution operators can be integrated into deep learning layers. This research proposes a new layer based on a second-order 1D Volterra series expansion using the TensorFlow environment. To develop the Volt1D, we first analyzed a linear convolutional layer's performance on a human speech dataset. The Volterra series has been particularly successful in speech recognition, as it allows for modeling the nonlinear dynamics of the human vocal tract. Volt1D allowed us to capture higher-order nonlinearities in the system, significantly improving the model's accuracy. To validate the effectiveness of the Volt1D, we conducted extensive experiments on a dataset of the human speech command. Overall, our research demonstrates the potential of the Volt1D as a powerful tool for training speech recognition models.