Leveraging Landmark Acoustic Features in Cochlear Implant Signal Processing

The Lombard Effect (LE), defined as acoustic changes in speech production due to auditory feedback from the speaker’s acoustic environment, has been shown to improve intelligibility in normal hearing and some hearing-impaired listeners. However, LE has not been investigated specifically for cochlear implant systems and users. For CI users, signal processing strategies generate an electric representation of the acoustic signal which has enabled high speech understanding performance in quiet conditions but declines in the presence of naturalistic noisy environments. In this thesis, a range of signal processing approaches are proposed to improve electric stimulation by influencing channel selection based on salient speech features and to leverage the acoustic properties of Lombard speech as a means to improve intelligibility for CI users in difficult listening scenarios. Traditional ‘n’-of-‘m’ processing strategies utilize an energybased channel selection criteria to select ‘n’ out of ‘m’ available channels corresponding to the intracochlear electrode array. For speech-in-noise scenarios, noise-dominant channels may be selected at the expense of speech frequency-rich channels carrying important phonetic cues. For quiet listening scenarios, low-level consonant energy may be overshadowed by higher-level channels. To overcome these challenges, a formant-based channel selection scheme is used to determine the effect on channel selection and speech intelligibility. This approach is hypothesized to illicit minor yet strategic changes in the electric representation of speech to include formant frequency information in the presence of noise. Second, two apriori compression functions are proposed to increase the intensity of formants and consonant segments. Lastly, three proposed speech modification strategies inspired by Lombard speech are introduced. The ability of CI listeners to perceive LE, the benefits of LE perturbation of neutral input speech, and the effect of semantics on LE perturbation are all addressed. These perturbation approaches are hypothesized to illicit large, meaningful changes in the electric representation by altering the following changes in the speech structure: intensity, first formants and second formant location/amplitude/bandwidth, long-term average spectrum, fundamental frequency, and individual phoneme class duration. We demonstrate the implications of CI listeners to leverage salient features Lombard speech and examine the feasibility of Lombard perturbation to improve speech understanding for CI users.