Towards objective evaluation of audio time-scale modification methods

Abstract

The need for high-quality time-scale modification of audio is increasing, as media streaming services are providing new related functionalities to their users. The main goal of a time-stretching method is to preserve the pitch and the subjective quality of the different components of the audio signal, namely transients, noise, and tonal components. Many solutions have been proposed throughout the years, with various results depending on the kind of processed audio input. This paper introduces an evaluation method for audio time-scaling algorithms based on a recent fuzzy time-frequency decomposition, which reveals the energy of the tonal, transient, and noise components in the original and stretched sounds. From the energy curves, typical impairments, such as transient smearing and the loss of tonality, can be observed. This analysis approach is compared with the subjective preferences of different techniques. This leads to suggestions for possible improvements of future algorithms. The ultimate goal is having an objective evaluation method which matches the subjective quality assessment.