Vowel identification models inject noise scaled to each user’s spectral modulation threshold

Evidence

The model estimates relevant features from electrodograms generated by the CI sound processor. Because we are modelling a vowel identification task it makes sense to extract features related to formant frequencies. Moreover, because we are analyzing the effect of enhancing spectral contrast it seems logical to use spectral contrast features between formants. In this study the number of features was limited to two formants and therefore the MPI model is two dimensional. Next, the MPI model adds noise to the features. The variance of the noise is set based on the individual abilities of a CI user to perceive the features extracted from the electrodogram. In our implementation we used the results of the SMT task to set the noise variance. The obtained jnd from the SMT task was scaled between 0.001 and 0.5. This number was applied as the variance of the Gaussian noise applied in the MPI model and for this reason is termed jnd noise in Fig. 1 . — Dijk, Pim van, Physiology, Psychoacoustics and…, 1831

Categorization

• QUESTION: How well does per-listener jnd-scaled noise in an MPI model actually predict real CI vowel-confusion patterns?—worth chasing the validation method later in the book.
• corroborated: 9 independent passages support this claim — Dijk, Pim van, Physiology, Psychoacoustics and…, 1831
• XREF: connects to: Internal noise, not feature extraction, explains individual variation in model vowel identification; Multi-dimensional phoneme identification models add noise scaled to individual perceptual acuity; Frequency importance functions weight speech audibility by band significance