Real-time speech/music classification with a hierarchical oblique decision tree

In the problem of classification of audio signals, the requirements of low-complexity, high-accuracy and short delay are crucial for some practical scenarios. This paper proposes a method of real-time speech/music classification with a hierarchical oblique decision tree. A set of discrimination features in frequency domain are selected together with a proposed simple harmonic structure stability feature, which is based on a rough estimation of the harmonic structure. A feature subset selection tool is used to select a subset of short and long term features to feed into a hierarchical oblique decision tree classifier. The method is evaluated and compared with the open loop selection mode in AMR-WB+. Experiments show the proposed approach gives a better performance (98.3%) compared to other prevailing approaches. In particular, it comes with promising short delay of 10 ms and low complexity of 1 wmops.