Objectively Choosing Spectrogram Parameters to Classify Environmental Noises

摘要

Spectrograms are commonly used to visualize, analyze, and classify audio signals in the same way that social media companies (e.g., Google, Facebook, Yahoo) use images to classify or tag people in photos. A problem unique to using spectrograms to classify acoustic signals is that the user must choose the spectrogram input-parameters, which may affect the accuracy of the resulting classifier. While the spectrogram - in its simplest form - only has three input-parameters, each parameter has a large number of possible values it can take, resulting in a nearly infinite number of combinations and unique spectrograms. The three input-parameters include the window-type, window-size, and percent-overlap-between-windows. The process of choosing spectrogram parameters, however, is often glossed over in the literature, and there is typically little guidance on how to make this, often, subjective choice. We hypothesize that the choice of spectrogram input-parameters will affect the spectrogram output or features that in turn will affect the performance of the acoustic classifiers. To test this hypothesis, we use Matlab's built-in spectrogram function, a support-vector-machine classifier, a labeled (i.e., human classified) environmental noise dataset, and randomly sample the spectrogram input-parameter space to objectively choose the spectrogram input-parameters. We find that the random sampling procedure is a useful way of choosing the spectrogram input-parameters, and finding the spectrogram features that are the most important for classifying environment noises. The environmental noises used in this study include the noise from air conditioners, car horns, children playing, dogs barking, drilling, engine idling, gunshots, jackhammers, sirens, and street music.