Multiple single sound events of very different characteristics might coincide in a given space and time, thus composing complex audio scenes. In that context, defining signal features capable of effectively analyzing the holistic audio scenes is a challenging task. This paper introduces a set of features that consider the temporal, spectral and perceptual characteristics of the audio scene signals. Specifically, the features are obtained from the autocorrelation function of band-pass signals computed after applying a Mel filter bank. The so-called Narrow-Band Autocorrelation (NB-ACF) features are compared to state-of-the-art signal features on a corpus of 4 hours composed of 15 audio scenes. Regardless of the learning algorithm employed, the NB-ACF attains the highest averaged recognition rates: 2.3 % higher than Mel Frequency Cepstral Coefficients and 5.6 % higher than Discrete Wavelet Coefficients.