Identification of environmental sounds with varying spectral resolution

This study investigated the identification of familiar environmental sounds with varying spectral resolution to establish (1) the number of frequency channels needed to perceive a large heterogeneous set of familiar environmental sounds, (2) the role of cross-channel asynchrony in identification performance, and (3) the acoustic correlates of the spectral resolution required for identification.

In experiment 1, 60 normal-hearing listeners identified environmental sounds in a 60-alternative closed--set response task as a function of six spectral resolution conditions (i.e., 2, 4, 8, 16, 24, and 32 frequency channels) obtained with an envelope-vocoder. In experiment 2, identification accuracy for varying amounts of cross-channel asynchrony was determined for sounds with preserved and degraded fine spectral structure in 10 normal-hearing listeners. Experiment 3 examined identification performance of 72 listeners across six spectral resolution conditions as in experiment 1, but using three different signal processing methods designed to minimize cross-channel asynchrony across channels. Follow-up acoustic and discriminant analyses were carried out to identify parameters that can distinguish environmental sounds based on required spectral resolution.

Identification accuracy tended to improve with increasing spectral resolution reaching the maximum of 76%. However, in experiment 1, performance did not change significantly beyond eight channels, whereas identification accuracy of some sounds declined with increasing spectral resolution. In experiment 2, increases in cross-channel asynchrony for sounds with preserved fine spectra had a small, but significant negative effect on identification. However, minimizing the amount of asynchrony had no significant effect on the overall identification of spectrally degraded sounds in experiment 3. Acoustic analysis indicated several spectral and temporal measures that differed significantly between sounds that required eight or fewer channels and those that required 16 or more channels for 70% correct identification. Discriminant analysis revealed that the sounds could be classified into high- and low-required spectral resolution groups with 83% accuracy based on only two acoustic parameters: the number of bursts in the envelope and the standard deviation of spectral centroid velocity.

Increasing spectral resolution generally had a positive effect on identification of familiar environmental sounds. However, across conditions performance accuracy remained well-below that of control stimuli with preserved fine spectra, despite becoming asymptotic above eight channels. Cross-channel asynchrony introduced during vocoder processing, although detrimental for some sounds, was not a major factor that prevented further improvement in overall accuracy. A spectral resolution greater than 32 channels, along with additional fine spectral and temporal information may be required for identification of a number of environmental sounds. This study provides a preliminary basis for optimizing environmental sound perception by cochlear implant users by highlighting the role of several acoustic factors important for environmental sound identification.