1	Identifying the Sources of Environmental Sounds with a Varying Number of Spectral Channels Valeriy Shafiro, James J. Jenkins, Winifred Strange City University of New York, Graduate Center
2	Research Questions and Significance What is the minimal spectral resolution required for source identification? Can environmental sounds be grouped based on the required spectral resolution? A framework for exploring acoustic parameters and cognitive factors in source identification. Knowledge base for cochlear implant design and performance predictions.
3	Method: Signal Processing Sounds filtered with 6th order Butterworth filters (overlapping at - 3dB) into N log spaced frequency bands (300 to 5500 Hz). Envelopes were obtained for each band via half-wave rectification followed by lowpass filtering at 160Hz. Each envelope was excited with white noise and filtered using original filter settings. Resulting individual passband signals were combined.
4	Method: Example of spectral variation
5	"Stimuli and design" Stimuli and design: 60 familiar environmental sounds (Mean correct 97%); Six channel conditions: 2, 4, 8, 16, 24, 32. Latin square design - no listener heard the same type of sound in more than one channel condition. Undistorted sounds were presented at the end. Task: 60 alternative closed-set forced choice . Listeners: 60 listeners (40 f, 20 m. / 37 NE, 23 NN) All passed a hearing screening at 25 dB HL
6	Results: Identification accuracy
7	Why does performance asymptote? Identification accuracy of 19 sounds decreased by > 1SD (30%) as the spectral resolution increased. Most affected sounds were broadband, temporally patterned (e.g., ‘clapping’, ‘horse trotting’) Decreased accuracy was due to temporal distortions (unequal group delays?). 41 remaining sounds: 30 improved, and 11 were asymptotic (within the 30% error margin).
8	Identification accuracy re-examined
9	Grouping of sounds by number of channels (at 70%)
10	Cross channel correlations Accuracy (Spearman ranks) Response Frequency (Pearson)
11	Conclusions Increasing spectral resolution by raising the number of frequency channels improves source identification. The rate of improvement in source identification performance decreases beyond 16 channels. Spectral asynchrony caused by filter group delays adversely affects source identification performance with some sounds. The perceptual properties of individual sounds can be differentiated by the amount of spectral detail required for source identification. Different degrees of spectral resolution elicits different response patterns.
12
13
14	Average group delays per band