Spectro-temporal modulations (STMs) of the sound convey timbre and rhythm information so that they are intuitively useful for automatic music genre classification. The STMs are usually extracted from a time-frequency representation of the acoustic signal. In this paper, we investigate the efficacy of two kinds of STM features, the Gabor features and the rate-scale (RS) features, selectively extracted from various time-frequency representations, including the short-time Fourier transform (STFT) spectrogram, the constant-Q transform (CQT) spectrogram and the auditory (AUD) spectrogram, in recognizing the music genre. In our system, the dictionary learning and sparse coding techniques are adopted for training the support vector machine (SVM) classifier. Both spectral-type features and modulation-type features are used to test the system. Experiment results show that the RS features extracted from the log. magnituded CQT spectrogram produce the highest recognition rate in classifying the music genre.