NOTE: This repository hasn't been updated since October 2020 due to confidentiality of associated institutions.
The implementation of a typical neural network for Emergency Sounds Identification includes three steps: training, validation and testing. In the training stage, the extracted MFCC features of training dataset are provided as input to estimate the weight parameters in a neural network. In the validation phase, we use the validation dataset to evaluate the performance of trained model, and through this basis to tune the parameters to achieve the best one. Lastly, in the testing phase we pass the unknown audio signal to the fully-trained network to obtain corresponding output, the output with maximum response value is selected as identified result.
In time domain, the processing of signals on computers is more complicated, so we will analyze the features in the frequency domain instead. In addition, since the sound data is large in both time and frequency domain, in order to reduce the amount of data to be processed, the short-term features must be extracted to represent the audio segments. Currently, the widely used features in audio signal processing are Perceptual Linear Prediction (PLP) coefficients, Linear Prediction Coefficients (LPC) and Mel-Frequency Cepstral Coefficients (MFCCs). In general, MFCCs are used as phonological feature parameters in speaker recognition application. However, it has been shown viable in emergency signal recognition system effectively, and hence we use MFCCs in this work as well. For every 3s of a sound signal, 5 segments each and ultimately 13- dimensional MFCCs are computed from their spectral representations known as spectograms, derived with the help of the librosa library.
A dense neural network implementation is done using Keras. 'Relu' activation along with a 30% dropout in hidden layers makes the model prevent overfitting. 'Softmax' activation is used in the final three-way classifier output layer. 'Adam' Optimizer helps in learning the data (MFCC vectors) by tuning weights of the hidden layers.
numpy 1.18.5tensorflow 2.3.1librosa 0.8.0