Speech Perception in Noise for Cochlear Implant Users

A cochlear implant (CI) is one of the most successful and effective ways to treat severe to profound hearing loss problems. According to the FDA, more than 300,000 profoundly deaf people have been implanted with a CI and this number is growing quite rapidly. It is so fascinating that an electrical instrument can enrich people's life with sound, who were previously living in a world of (almost) complete silence. The effectiveness of CIs is dependent upon many factors but in general, a satisfactory speech understanding performance in favorable conditions is reported where most CI users are able to communicate in relatively quiet listening environments. However, speech understanding performance in noisy environments is still significantly poorer compared to the normal hearing (NH) listeners. Most CI recipients typically require about 5-25 dB higher Signal-to-Noise Ratio (SNR) compared to NH listeners in order to achieve similar speech understanding. CI recipients have an impaired auditory system with often limited neural survival (depending upon the etiology and duration of deafness) which restricts the transmission of both fine temporal and spectral information to the auditory cortex. Furthermore, the electrical stimulation bottleneck of the auditory nerve further restricts the transmission of independent information due to its limited dynamic range and broad excitation pattern. In addition, speech coding strategies may not necessarily encode the fine temporal and spectral information in an optimal way to be perceived by the CI users. To narrow the performance gap between the NH and CI users, signal processing needs to be optimized specifically for CI users to improve the transmission of perceivable independent information to the auditory system. Improving the speech perception in noise is one of the most important goals of research in CIs and is also the topic of this PhD thesis. Chapter 1 introduces the basics of human auditory system and coding mechanisms in state of the art CI systems along with the bottlenecks of electrical hearing. In response to the incoming sound the speech coding strategy in a CI determines the amplitude, the width, the timing and the number and location of the electrodes/channels to be stimulated. In noisy conditions however the amplitude and channel selection does not remain ideal. In chapter 2 several research questions are addressed to find out the most important factors that limit the intelligibility of the CI processed speech in noisy environments. The objective of the first study (chapter 2) is to characterize and understand the effect of noise on the performance of CI users and to understand what the similarities and differences are between the noise reduction requirements of CI vs. NH listeners. To serve this purpose, the influence of low frequency ON/OFF modulations, target speech distortions and channel selection issues on speech intelligibility are investigated. It is found that the degradation in speech understanding performance is mainly due to the distortions in low frequency ON/OFF modulations and channel selection errors. Furthermore, an ideal channel selection strategy that can avoid the maxima selection errors improves speech intelligibility even without suppressing the noise in the target speech gaps. The study also highlights the differences in noise reduction requirements of CI users compared to NH listeners. It is shown that CI users can accept significantly lower levels of noise in the target speech gaps but can tolerate high levels of distortions in the target speech segments. This allows more aggressive noise reduction for CIs.Intelligibility improvements provided by the single channel noise reduction algorithms for CI remains small due to the estimation errors in the parameters for filtering and therefore cannot bridge the speech understanding performance gap between NH and CI recipients. In recent years significant emphasis has been put on multi-channel noise reduction algorithms which not only have two or more microphone signals at their disposal to generate an output, but also make use of the head shadow effect in the binaural case. The study described in chapter 3 which evaluates a binaural noise reduction algorithm based on the phase error distribution provides significant improvement in speech understanding for both NH and CI users. This study further confirms the aggressive noise reduction requirements of CI users. The study reveals that NH listeners have a strong preference for the soft masks compared to binary masks whereas CI subjects show no preference.Chapter 4 and 5 of this thesis present two novel concepts to improve state of the art coding strategies and speech perception in noise. In chapter 4 the emphasis is on the input processing and the study investigates whether a SNR based Channel Selection (SCS) criterion can provide speech perception benefit over the channel selection based on maximum energy. The maxima selection in ACE can select masker dominated channels in noisy conditions which can be avoided by selecting the channels on the basis of the SNR estimates. The results show that the noise is significantly less annoying with the new channel selection criterion. However, this improvement was not translated into significantly better speech understanding scores. The possible reasons for no improvement in speech understanding scores are discussed and it is concluded that suppression of the noise may be necessary to improve speech understanding along with the SNR based channel selection.In Chapter 5 the emphasis is on improving the spectral representation of the acoustic signal in the electrical stimulation. This study presents and evaluates the concept of complete signal spectrum stimulation and correlation maximization concepts to improve the spectral shape representation of the electrical stimulation and therefore the speech understanding scores in noisy conditions. ACE stimulates only part of the signal spectrum due to the n-of-m maxima selection and therefore does not follow the principle of superposition, which results in a decrease of speech understanding performance in conditions where the channel frequency response is non-ideal. The proposed Frequency response Robust ACE (FRACE) uses the concept of complete spectrum stimulation together with the ACE channel selection to make it more robust against the channel frequency response. The results show that FRACE provides significant improvement in speech understanding compared to ACE in non-ideal channel response conditions. The complete signal stimulation concepts in FRACE tries to recreate the complete spectral shape of the acoustic signal in the electrical stimulation but does not take into account the spread of excitation or refractory properties of the neural response. The Correlation Maximization (CM) strategy however, models the excitation pattern in the cochlea by explicitly accounting for the excitation spread and refractory behavior of the neural response and tries to maximize the correlation between the desired and the modeled excitation pattern. It is shown that the relative strength of different speech formants is not preserved in the modeled excitation pattern due to the combined effect of maxima selection and excitation spread. The perceptual evaluations in speech shaped and party noise reveal a significant benefit of the CM Channel Selection (CMCS) strategy over ACE. These results are encouraging and suggest that the concept of improving the spectral shape representation by taking the spread of excitation and refractory properties into account is promising for the improvement of speech understanding in noisy environments. Overall the studies described in this thesis provide valuable new information regarding the noise reduction requirements of the CI users and provide an important step towards the development of an enhanced new generation coding strategy.status: publishe