Unsupervised Neural Network approach for the Identification of Anomaly in Speech Signal from Spectrogram Images

Abstract

In this paper, anomaly identification in speech signal is carried out using an unsupervised Neural Network approach. Input audio signal is divided into small segments of information. Initial segment of the samples are used in the training process. The later portion of the sample segments are used in the test process. The features (Spectral Roll-off, Spectral Centroid, Mel Frequency Cepstral Coefficient (MFFC), Pitch (PH) and Energy Density (ED)) are extracted from the input data. The extracted 1D features are converted into spectrogram images. Then, the images are fed as an input to Neural Network for the prediction of feature values. If three or more feature value does not exceed the threshold value of 75% then the input signal is considered to be free from anomaly. The proposed model results in an accuracy of 97.50% in the detection of anomaly in input speech signal.