IEEE Project Abstract

Deep neural networks (DNNs) have been provento be powerful models for acoustic scene classification tasks.State-of-the-art DNNs have millions of connections and arecomputationally intensive, making them difficult to deploy onsystems with limited resources. With a focus on acoustic sceneclassification, we describe a new learnable module, the simulatedFourier transform module, which allows deep neural networksto implement the discrete Fourier transform (DFT) operation8x faster on a GPU. We frame the signal processing procedure as an adaptive machine learning problem and introducelearnable parameters in the module to facilitate fast adaptationfor the complex and variable acoustic signal. This modulegives neural networks the ability to model audio signals fromraw waveforms, without extra FFT and filter bank patches.Then we use the temporal transformer module, which has beenpreviously published, to alleviate the information loss caused bythe simulated Fourier transform module. These techniques canbe integrated into an existing fully connected neural network(FCNN), convolutional neural network (CNN) or recurrent neural network (RNN) model. We evaluate the proposed strategyusing four acoustic scene datasets (LITIS Rouen, DCASE2016,DCASE2017, DCASE2018) as target tasks. We show that theproposed approach significantly outperforms the vanilla FCNN,CNN and RNN approach on both efficiency and performance.For instance, the proposed approach can reduce inference timeby 8x while reducing the classification error on LITIS Rouendataset from 3.21% to 1.81%.