Using Machine Learning to Mitigate Reverberation Effects in Cochlear Implants

 描述(由申请人提供)：人工耳蜗(CI)为全球超过200,000名接受者提供听力[NIDCD，2011#637}。这些设备成功地在安静的听力条件下提供了高水平的言语理解；然而，更具挑战性的条件使CI接受者的言语理解能力下降的程度比正常听力听者要大得多(Kokkinakis，2011#673；Nelson，2003#302)。CI听者尤其受到混响条件的影响，即使是少量的混响也会比大量的稳态噪音更大程度地降低理解能力[Hazrati，2012#674}。因此，减轻混响影响的方法有可能极大地提高CI用户的生活质量。以前解决人工耳蜗混响中的语音问题的尝试未能实时实现。我们的初步结果表明，成功的重叠掩蔽的缓解可以导致语音识别的实质性改善，即使自我掩蔽没有得到缓解，我们已经设计了一种可以实时实现的方法。在提议的努力中，我们将首先在我们成功的初步努力的基础上改进用于检测混响的分类器。接下来，我们将评估缓解算法，首先是在正常听力听众中，然后是在植入人工耳蜗的听众中。最后，我们将实时实现该算法，并对其进行再次测试。

 DESCRIPTION (provided by applicant): Cochlear implants (CIs) provide hearing for over 200,000 recipients worldwide {NIDCD, 2011 #637}. These devices successfully provide high levels of speech understanding in quiet listening conditions; however, more challenging conditions degrade speech comprehension for CI recipients to a much greater degree than for normal hearing listeners {Kokkinakis, 2011 #673;Nelson, 2003 #302}. CI listeners are especially affected by reverberant conditions with even a small level of reverberation degrading comprehension to a greater degree than a large amount of steady-state noise {Hazrati, 2012 #674}. Thus, a method that mitigates the effects of reverberation has the potential to greatly improve the quality of life for CI users. Previous attempts to solve the problem of speech in reverberation for cochlear implants have not been able to be implemented in real time. Our preliminary results suggest that successful mitigation of overlap masking can result in a substantial improvement in speech recognition even if self-masking is not mitigated and we have devised an approach that can be implemented in real time. In the proposed effort, we will first improve the classifier to detect reverberation based on our successful preliminary efforts. Next we will assess the mitigation algorithm, first in normal hearing listeners and then in listeners with cochlear implants. Finally, we will implement the algorithm in real time and again test it.