Adaptive Feature Selection for Speech / Music Classification

In this paper, we propose a new system for classifying audio segments as speech or music. The proposed system improves classification accuracy, particularly in low signal-to-noise ratio (SNR) environments. The system selects the features with the highest classification accuracy that corresponds to the SNR value. The value of this features are compared to certain thresholds, which are also adapted to the SNR. Multi-expert method of combining the features to improve classification accuracy is implemented. A new feature, termed the variance of low-band energy ratio, is also introduced. This feature produces large improvements in classification accuracy at low SNR. Performance of the proposed system is evaluated for different SNR using a library of speech and music audio segments. Using one-second segments it is shown that the proposed system can enhance the classification accuracy by 22% at SNR=-15 dB, and obtain classification accuracy of 90.3% at SNR=0 dB