In this paper, the noise-robustness of a recently proposed fast Fourier transform (FFT)-based auditory spectrum (FFT-AS) is further evaluated through speech/music/noise classification experiments wherein mismatched test cases are considered. The features obtained from the FFT-AS show more robust performance as compared to the conventional mel-frequency cepstral coefficient (MFCC) features. To further explore the FFT-AS from a perspective of practical audio classification, an audio classification algorithm using features derived from the FFT-AS is implemented on the floating-point DSP platform TMS320C6713. Through various optimization approaches, a significant reduction in the computational complexity is achieved wherein the implemented system demonstrates the ability to classify among speech, music and noise under the constraint of real-time processing.
展开▼