Amplitude mapping and phoneme recognition in cochlear implant listeners

Speech and other environmental sounds must be compressed to accommodate the small electric dynamic range in cochlear implant listeners. The objective of this paper is to study whether and how amplitude compression and dynamic range reduction affect phoneme recognition in quiet and in noise for cochlear implant listeners.

Four implant listeners using the Nucleus-22 SPEAK speech processor participated in this study. The amount of compression was varied by manipulating the Q-value in the SPEAK processor. The size of the dynamic range was systematically reduced by increasing the threshold level and decreasing the comfortable level in the processor. Both female- and male-talker vowel and consonant materials were used to evaluate speech recognition performance in quiet and in noise. Speech-spectrum-shaped noise was mixed with the speech signal and presented continuously to the speech processor through a direct electric connection. Signal to noise ratios were changed over a 30 to 40 dB range, within which phoneme recognition increased from chance to asymptotic performance. Phoneme recognition scores were obtained as the number of active electrodes was reduced from 20 to 10 to 4. For purposes of comparison, phoneme recognition data also were collected in four normal-hearing listeners under comparable laboratory conditions.

In both quiet and noise, the amount of amplitude compression did not significantly affect phoneme recognition. The reduction of dynamic range marginally affected phoneme recognition in quiet, but significantly degraded phoneme recognition in noise. Generally, the 20- and 10-electrode processors produced similar performance, whereas the 4-electrode processor produced significantly poorer performance. Compared with normal-hearing listeners, cochlear-implant listeners required higher signal to noise ratios to achieve comparable recognition performance and produced significantly lower recognition scores at the same signal to noise ratios.

The amount of amplitude compression does not significantly affect phoneme recognition, whereas reducing dynamic range significantly lowers phoneme recognition, particularly in noise and for vowels. Because the SPEAK processor extracts mostly spectral peaks, the present conclusions may not be applied to other types of processors extracting temporal envelope cues. The present results also suggest that more than four electrodes are required to optimize speech recognition in multiple-talker and noise conditions. A significant performance gap in speech recognition still remains between cochlear implant and normal-hearing listeners at the same signal to noise ratios. Improved cochlear implant designs and fitting procedures are required to narrow and, hopefully, close this performance gap.