Comparative studies on machine learning for paralinguistic signal compression and classification

In this paper, we focus on various compression and classification algorithms for three different paralinguistic signal classification tasks. These tasks are quite difficult for humans because the sound information from such signals is difficult to distinguish. Therefore, when machine learning techniques are applied to analyze paralinguistic signals, several different aspects of speech-related information, such as prosody, energy, and cepstral information, are usually considered for feature extraction. However, when the size of the training corpus is not sufficiently large, it is extremely difficult to directly apply machine learning to classify such signals due to their high feature dimensions; this problem is also known as the curse of dimensionality. This paper proposes to address this limitation by means of feature compression. First, we present experimental results obtained by using various compression algorithms to compress signals to eliminate redundancy of the signal features. We observe that compared with the original features, the compressed signal features still provide a comparable ability to distinguish the signals, especially when using a fully connected neural network classifier. Second, we calculate the output distribution of the F 1-score for each emotion in the speech emotion recognition problem and show that the fully connected neural network classifier performs more stably than other classical methods.