Novel Features for Effective Speech and Music Discrimination

omer La. tudent.unsw.edu.au 1 ambi,(a,unsw.edu.aU tacmaeJIbD C u ienesLdni Abstract- Speech and music discrimination has gained much databases is still not mature. A speech/music discriminator popularity in recent years for efficient coding and automatic can help to move towards this goal. retrieval of multimedia sources and Automated Speech 4) Automatic channel selection of FM radio broadcast. Recognition (ASR). Two novel features that can be concatenated This can be used to look for music channels on radio with Mel frequency cepstral coefficients are presented in this broadcast or to switch channels when commercials appear. paper: Delta Cepstral Energy (DCE) and Power Spectrum Deviation (PSDev). Employing a Gaussian mixture model for Significant research has been conducted in the past few classification as a back-end to the system, a significant years into effective audio indexing. Zero crossing count [10], improvement in the error rate was found using these features. Cepstral and spectral features [2, 11], low frequency features The effects of different musical instruments on error rates were [5, 6] and entropy and dynamism features [1] have been used also analyzed. Low frequency musical instruments like piano for classification. Combination of a few features has also and electric bass guitar were found to be more difficult to been analyzed [12]. Nearest Neighbor, Neural Networks, discriminate from speech, however, the proposed features are Gaussian Mixture Model (GMM), Hidden Markov Model also able to reduce such errors significantly. (HMM), Nearest Feature Line, Adaboost and Support Vector Keywords- audio indexing, speech recognition, multimedia Machine (SVM) have been used as the back end. Due to the coding, Mel frequency cepstral coefficients, Gaussian mixture lack of a standard database for audio indexing systems, models. results are not easily comparable, however cepstral features, such as the Mel frequency cepstral coefficients (MFCCs