Recognition of Transient Environmental Sounds Based on Temporal and Frequency Features

Abstract

Environmental sound recognition (ESR) refers to the recognition of all sounds other than the human voice or musical sounds. Typical ESR methods utilize spectral information and variation within it with respect to time. However, in the case of transient sounds, spectral information is insufficient because only an average quantity of a given signal within a time period can be recognized. In this study, the waveform of sound signals and their spectrum were analyzed visually to extract temporal characteristics of the sound more directly. Based on the observations, features such as the initial rise time, duration, and smoothness of the sound signal; the distribution and smoothness of the spectrum; the clarity of the sustaining sound components; and the number and interval of collisions in chattering were proposed. Experimental feature values were obtained for eight transient environmental sounds, and the distributions of the values were evaluated. A recognition experiment was conducted on 11 transient sounds. The Mel-frequency cepstral coefficient (MFCC) was selected as reference. A support vector machine was adopted as the classification algorithm. The recognition rates obtained from the MFCC were below 50% for five of the 11 sounds, and the overall recognition rate was 69%. In contrast, the recognition rates obtained using the proposed features were above 50% for all sounds, and the overall rate was 86%.