Contribution of consonant landmarks to speech recognition in simulated acoustic-electric hearing.

OBJECTIVES: The purpose of this study is to assess the contribution of information provided by obstruent consonants (e.g., stops and fricatives) to speech intelligibility in simulated acoustic-electric hearing. As a secondary objective, this study examines the performance of an objective measure that can potentially be used for predicting the intelligibility of vocoded speech.
DESIGN: Noise-corrupted sentences are used in experiment 1 in which the noise-corrupted obstruent consonants are replaced with clean obstruent consonants, while leaving the sonorant sounds (vowels, semivowels, and nasals) corrupted. In one condition, listeners have only access to the low-frequency (<600 Hz) acoustic portion of the clean consonant spectra, in other condition, listeners have only access to the higher frequency (>600 Hz) portion (vocoded) of the clean consonant spectra, and in the third condition, they have access to both. In experiment 2, we investigate a speech-coding strategy that selectively attenuates the low-frequency portion of the consonant spectra while leaving the vocoded portion corrupted by noise. Finally, using the data collected from experiments 1 and 2, we evaluate the performance of an objective measure in terms of predicting intelligibility of vocoded speech. This measure was originally designed to predict speech quality and has never been evaluated with vocoded speech.
RESULTS: Significant improvements (about 30 percentage points) in intelligibility were noted in experiment 1 in steady and two-talker masker conditions when the listeners had access to the clean obstruent consonants in both the acoustic and the vocoded portions of the spectrum. The improvement was more evident in the low signal to noise ratio levels (-5 and 0 dB). Further analysis indicated that it was access to the vocoded portion of the consonant spectra, rather than access to the low-frequency acoustic portion of the consonant spectra that contributed the most to the large improvements in performance. In experiment 2, a small (14 percentage points) but statistically significant improvement in performance was obtained at 0 dB signal to noise ratio (steady masker) when the obstruent consonants were selectively attenuated in the low-frequency acoustic portion alone (the vocoded portion was left noise corrupted). The examined objective measure predicted with a relatively high correlation (r = 0.92 to 0.94) [corrected] the intelligibility of vocoded speech improved in both steady and two-talker masking conditions.
CONCLUSIONS: Providing access to the clean obstruent spectra can yield substantial improvements in intelligibility relative to the simulated acoustic-electric condition. Much of this improvement can be attributed to the listeners having access to the clean vocoded portion of the obstruent consonants. The large contribution of obstruent consonants in speech recognition in simulated acoustic-electric hearing stems from the fact that these consonants provide reliable acoustic landmarks which in turn enable listener to integrate effectively pieces of the message glimpsed over temporal gaps into one coherent speech stream. It is argued that these landmarks are smeared in existing cochlear implant systems, including the bimodal systems, owing to envelope compression, and the fact that the obstruent consonants are probably the first to be masked by background noise. Overall, the outcomes from this study suggest that the obstruent consonants need to be treated differently for improved speech recognition in noise.