Speech/Music Discrimination in Time Domain

Classifying an audio signal into either speech or music has been of continuous interest to researchers. The literature shows a variety of methods used. Each method investigates the usage of a combination of features and further analyses them in time/frequency domain. Some methods combine both domains; i.e. time-frequency analysis. Evidently, time-domain analysis is more suitable for real-time applications. In spite of being heavily researched, features in the time domain still have gaps and questions to be answered. Among these is a question that complements the previous works on zero crossing rate (ZCR). The zero crossing is not the only event that exists in a sampled audio signal. Other events do exist. The question is: what are the other possible events that could carry information towards the discrimination between speech and music? And how do they perform in comparison to ZCR? Relevance The methods so far vary in terms of their performance both time-wise and computation-wise. The former is important for real-time applications. The latter is important for cheaper implementation. Most researchers tend to work in one of two ways: i) Group a number of features into one vector to be used for discrimination ii) Use a different classifier that has not been used yet for audio classification; e.g. Neural Network, Fuzzy Classifier, Nearest Feature Line, etc. The approach in this project is slightly different. The hypothesis is