Simulated cochlear-implant processing results in major loss of acoustic information regarding differences in talkers’ voice qualities

Abstract

The ability to recognize different voice qualities is essential for good talker identification; yet, little is known about how well voice quality cues of talkers are transmitted through the degraded speech signals delivered by cochlear implants (CIs). This study examined how CI speech filtering affects acoustic distinctiveness of individuals with and without voice disorder. Sustained /a/ vowels uttered by speakers with normal or disordered voice were processed using 4, 8, 12, 16, 22, and 32 channel noise-vocoders. The effect of CI processing on the distinctiveness of talkers with normal and disordered voices was measured using the Mahalanobis distance measure on Mel-frequency cepstral coefficients derived from samples across these groups. The analysis confirmed that CI vocoding dramatically degrades acoustic cues in frequency sub-bands that signal abnormal voicing behavior. Remarkable spectral degradation was observed in low- ( 12 kHz) bands for simulated conditions compared with unprocessed signals. These findings indicate that CI users likely have almost no ability to distinguish talkers differing in voice qualities. These results highlight challenges that CI users face for recognizing talkers differing in voice quality, due to these users’ lack of access to fine-grained spectro-temporal details in voices.