Speech enhancement for Cochlear implants

Approximately 360 million people in the world suffer from a disabling hearing loss. A cochlear implant (CI) is a successful treatment for people with profound sensorineural hearing loss and evokes a hearing sensation by direct electrical stimulation of the auditory nerve fibers (ANFs). Tremendous technological advances in the last decade make it possible that a lot of CI users are able to understand speech very well in quiet listening conditions. However, the speech intelligibility (SI) performance decreases rapidly in adverse listening conditions with interfering background sounds. Speech enhancement algorithms are developed to try to overcome the deficit of poor speech perception in speech in noise scenarios.There are two general approaches for speech in noise enhancement. One approach is called speech enhancement strategies that try to improve SI by an intentional emphasis of speech cues important for speech perception. Another approach are the noise reduction (NR) algorithm that try to reduce as much noise as possible from the noisy mixture of the target signal and the interfering background sound to improve intelligibility. This dissertation deals with the development and analysis of both kinds of speech in noise enhancement strategies for the application in CIs.The first study focused on the development of a new physiologically motivated speech enhancement strategy called enhanced envelope (EE) strategy that mimics the adaptation characteristics of the hair-cell-to-neuron synapse at the onsets of the speech envelope. In a CI, the synapse is bypassed by the direct electrical stimulation of the ANFs. The potential and feasibility was assessed with normal-hearing (NH) subjects listening to noise vocoder CI simulations. A significant speech reception threshold (SRT) improvement was obtained in stationary speech-shaped noise (SSN) and with a competing talker when the onsets were enhanced with a priori knowledge of the target signal. The extension to real-time applications was investigated with a Wiener filter NR algorithm that was used in the onset extraction stage.In the second study, the potential of the EE strategy for the application in CI was investigated with CI users. A significant improvement in speech recognition was obtained in stationary SSN, with an interfering talker, and in a stop consonant recognition task. The loudness perception was not influenced by the processing. The results suggest that an enhancement of the onset cues of the speech envelope can lead to a better parsing of the auditory scene and, therefore, support the segregation of the desired speaker from the background sounds.In the third study, the potential of ideal time-frequency masking approaches in single-channel NR systems for SI and speech quality improvement was investigated with NH listeners. The ideal Wiener filter (IWF) as a representative of a mask with continuous weight was compared to the ideal binary mask (IBM) as a representative of a binary mask. The comparison was done in terms of SI under ideal and with corrupted parameter estimates for the mask pattern derivation and in terms of speech quality. It was shown that the IWF vastly outperformed in all tasks the IBM when the frequency resolution was chosen similar to state-of-the art hearing aids (HAs) and CIs. The results suggest that the IWF should be preferred over a hard-decision approach in the application of auditory prostheses.In the fourth study, the IBM and IWF were systematically compared for the specific application in CI. It was shown that for CI users, the choice of mask pattern was not crucial and that there is no clear preference in terms of SI and speech quality in most conditions. In the fifth study, the application in HAs of the IBM and IWF was investigated with hearing-impaired (HI) listeners. Both mask patterns were not able to restore perfect SI. However, SI of the IWF processed signals was significantly higher than of IBM processed signals. In terms of speech quality, there was no significant preference for one of the NR strategies. The results underline that the development of speech enhancement algorithms have to take the needs of the respective target group into account because results obtained with NH listeners do not translate in all cases to results with CI and HA users.In multiple studies in this thesis, the potential of single-channel speech enhancement strategies was exploited. It was shown that envelope enhancement and NR algorithms have the potential to vastly improve SI of CI users in adverse listening conditions.status: publishe