Removing background talker noise for cochlear implant users

项目摘要/摘要 当听力受损的听众得到助听器(HA)或人工耳蜗(CI)的适当帮助时， 他们通常能够在安静的环境中舒适地保持对话。然而，在组中 环境，如大型家庭聚餐、餐厅或其他多人交谈的环境 同时，听力受损的听众在参与对话方面有很大的困难，而且经常 撤回或避免这种情况。因此，在HAS或CI中实现算法将非常有益 将背景说话者从信号中移除，以减少听障人士的听力负担 倾听并允许他们交谈，就像他们在一个安静的环境中一样。尽管HAS和CI经常 结合降噪算法，这些算法在背景杂乱无章的情况下是不有效的。这个 去除乱七八糟的问题涉及将语音与语音分开。因此，波谱特性 信号和噪声极其相似。 尽管有这些挑战，我们还是开发了一种非常有效的算法，名为SEDA来删除 背景胡言乱语。在iPhone上实现了SEDA的原型，并在10个CI用户上进行了评估。 在所有测试的信噪比(SNR)下，SEDA提高了对背景说话者的语音理解； 平均而言，学生对单词的理解能力提高了31个百分点。相比之下，最先进的 用于CI的降噪系统对于理解带有乱七八糟噪声的语音几乎没有好处。 CI制造商对我们的算法的成功概念验证表现出了极大的热情。 然而，在商业化之前，CI制造商希望降低计算能力 算法所需的。因为CI处理器最大限度地减少了计算处理，从而最大化电池 生活中，最大限度地减少SEDA所需的额外计算是很重要的。当使用SEDA作为前线时- 对于CI处理策略(就像我们的iPhone原型的情况)，所需的冗余 计算会导致计算量和延迟的增加。具体而言，SEDA将输入信号分解为 多个通道，去除背景杂乱无章，然后将它们重新组合成单个波形。这 然后，波形被馈入CI，CI再次将信号分解成多个通道。整合SEDA 进入信号处理链将节省计算处理，因为信号将只需要 分解一次，不需要重新组装。此外，尽管SEDA在 在噪音测试中，CI制造商强调了在更多情况下评估SEDA的重要性 真实的环境。 两个具体目标将解决CI制造商的商业化要求： 通过将SEDA集成到声音处理算法中来降低计算需求 在现实环境中评估SEDA。

PROJECT SUMMARY / ABSTRACT When hearing-impaired listeners are properly aided with a hearing aid (HA) or cochlear implant (CI), they are often able to comfortably maintain a conversation in quiet environments. However, in group environments, such as a large family dinner, restaurant, or other environment where multiple people are talking simultaneously, hearing-impaired listeners have great difficulty participating in conversations and frequently withdraw or avoid the situation. As such, it would be highly beneficial to implement an algorithm into HAs or CIs to remove background talkers (“babble”) from the signal to reduce listening effort for the hearing-impaired listener and allow them to converse as if they were in a quiet environment. Although HAs and CIs frequently incorporate noise reduction algorithms, these algorithms are not effective when the background is babble. The problem of removing babble involves segregating speech from speech. Hence, the spectral properties of the signal and noise are extremely similar. Despite these challenges, we developed an extremely effective algorithm named SEDA to remove background babble. A prototype of SEDA was implemented on an iPhone and evaluated on 10 CI users. SEDA improved understanding of speech with background talkers at all signal-to-noise ratios (SNRs) tested; on average, word understanding in babble improved by 31 percentage points. By contrast, the state-of-the-art noise reduction systems for CIs provide little to no benefit for understanding speech with babble noise. CI manufacturers have shown great enthusiasm about our successful proof-of-concept of our algorithm. Nevertheless, before commercialization, CI manufacturers want reductions in the computational power required for the algorithm. As CI processors minimize computational processing in order to maximize battery life, it is important to minimize the additional computations required by SEDA. When using SEDA as a front- end for a CI processing strategy (as is the case with our iPhone prototype), redundancy in the required calculations result in increased computations and latency. Specifically, SEDA decomposes the input signal into multiple channels, removes the background babble, and then reassembles them into a single waveform. This waveform is then fed into a CI which again decomposes the signal into multiple channels. Integrating SEDA into the signal processing chain will save computational processing as the signal would only need to be decomposed once and would not need to be reassembled. Additionally, although SEDA is highly successful in typical speech in noise tests, CI manufacturers emphasized the importance of evaluating SEDA in more realistic environments. Two specific aims will address the requirements for commercialization by the CI manufactures: reducing the computational requirements by integrating SEDA into a sound processing algorithm and evaluating SEDA in realistic environments.