Masking complex signals

In the past, psychoacoustic research on sound masking by the human ear mostly used simple sounds for the masker and probe signals (also called test signal). For instance, a narrow band of noise could serve as the masker, and a sinusoidal tone as the probe signal. Any observed degradation in detectability of the test signal when the masker is present is usually explained as the result of changes in the signal-to-noise ratio (SNR) in a bank of filters modelling the human ear, with overlapping frequency responses. As the difference between the frequencies of masker and probe increases, masking of the test signal decreases because the SNR in the filter processing the probe is less affected by leakage of signal energy from the filters processing the masker.

However, most if not all natural sounds are not simple but complex or compound sounds, i.e. they are made from frequency components or partials distributed over a specific bandwidth. Complex sound is often made of one basic sound and more or less natural harmonics. The filter bank model process each partial independently and predicts little or no interactions among them when the frequency gap becomes large enough. Nonetheless, there is plenty evidence that such an interaction exists. A harmonic template theory model, which has been proposed to account for such effects, holds that harmonically related sounds are processed in a special way by the human auditory system.

During development of objective quality measurement tools for audio signals, we noticed that some complex sounds with strong harmonic content that had been compressed to a low data rate by an audio codec contained quantization noise that was more obvious than predicted by the objective model. Thus began a program of research to assess and model the role harmonicity plays in masking complex sounds. In short, a high harmonic masker seems less effective than an inharmonic masker. Once fully understood, that phenomenon can be reproduced and included in psychoacoustic models to enhance the performance of codecs and to improve the accuracy of objective quality measurements techniques.

Modeling individual differences

Psychoacoustic models used in audio codecs or objective quality measurement tools are based on parameters in the scientific literature that reflect average performance of participants in psychoacoustic experiments. It is known, however, that wide individual differences exist in the population of listeners. Further, not all listeners are equal with respect to perceptual abilities. For example, some may be very good at frequency discrimination but not so at discriminating temporal gaps, while others may have good temporal discrimination and poor frequency discrimination.

Our experiments were intended to better understand the mechanisms that define or affect the different ranges of skills seen among listeners, particularly experts. Part of this effort is the development of tools and testing methods to measure more efficiently the properties of an individual auditory system. The purpose is to reduce time and efforts needed to characterize the auditory abilities of a relatively large population of listeners.