This study evaluated the performance of nine adaptive antifeedback algorithms. There were two goals: first, to identify objective procedures that are useful for evaluating these algorithms, and second, to identify strengths and weaknesses of existing algorithms.
The algorithms were evaluated in behind-the-ear implementations on the Knowles Electronics Manikin for Acoustic Research (KEMAR). Different acoustic conditions were created by placing a telephone handset or a hat on KEMAR. Electroacoustic techniques were devised to measure the following performance aspects of each algorithm: (1) additional gain made available before oscillation, (2) gain lost in specific frequency regions, (3) reduction of suboscillatory peaks in the frequency response, (4) speed of adaptation to changing acoustic conditions, and (5) robustness in the presence of tonal input signals.
For each measurement, performance varied widely across algorithms. No single algorithm was clearly superior or inferior to the others. Generally, the feedback cancellation algorithms were less likely to sacrifice gain in specific frequency regions and better at reducing suboscillatory peaks, whereas the algorithms that used noncancellation techniques were more tolerant of tonal input signals. For those algorithms equipped with special operational modes intended for music listening, the music mode improved the response to tonal inputs but sometimes sacrificed other performance aspects. Algorithms that required an acoustic measurement for initialization purposes tended to perform poorly in acoustic conditions dissimilar to the condition in which initialization was performed.
The objective methods devised for this study appear useful for evaluating the performance of adaptive antifeedback algorithms. Currently available algorithms demonstrate a wide range of performance, and further research is required to develop new algorithms that combine the best features of existing algorithms.