Relating the acoustic space of vowels to the perceptual space in cochlear implant simulations

In automatic speech recognition, speech features are measured and used as reference templates for machine learning and recognition. The differences between these features can be used to calculate relative acoustic distances between phonemes. However, when speech signals are spectrally degraded, as in electric hearing with cochlear implants, it is unclear whether these acoustic distances can predict speech recognition performance. The present study measured acoustic distances between spectrally degraded vowel tokens and investigated the relation between acoustic vowel space and perceptual vowel space. After processing vowel tokens using a cochlear implant simulation, the Mel-frequency cepstrum was extracted from each token; features were then time aligned and the weighted Euclidean distance was calculated between all tokens. Results demonstrate a significant correlation between vowel perception data and averaged acoustic distance between vowel tokens, for a variety of experimental conditions. These results suggest that acoustic distance between phonemes may well predict the recognition performance of spectrally degraded speech.