Comparison of ISFs and LSFs in Speech/Music Discrimination System

The immittance spectral frequencies (ISFs) is proposed as a new set of classification features and compared with the linear spectral frequencies (LSFs) applied in a frame-level wideband speech/music discrimination system. These two sets of features can be shared by the classifier and coding module to reduce the total computational complexity, making our classification system suitable for multi-mode audio coding applications. A performance assessment and comparison of the features are made. The experiment results show that the ISFs and LSFs have similar good performance when using full covariance matrices in classification models and the ISFs perform slightly better when using diagonal matrices. Their statistical differences for speech and music signals are also revealed.