A Real-Time Smartphone App for Unsupervised Noise Classification In Realistic Audio Environments

This paper presents a real-time unsupervised noise classifier smartphone app which is designed to operate in realistic audio environments. This app addresses the two limitations of a previously developed smartphone app for unsupervised noise classification. A voice activity detection is added to separate the presence of speech frames from noise frames and thus to lower misclassifications when operating in realistic audio environments. In addition, buffers are added to allow a stable operation of the noise classifier in the field. The unsupervised noise classification is achieved by fusing the decisions of two adaptive resonance theory unsupervised classifiers running in parallel. One classifier operates on subband features and the other operates on mel-frequency spectral coefficients. The results of field testing indicate the effectiveness of this unsupervised noise classifier app when used in realistic audio environments.