Commercial Readiness of a CI NR algorithm

项目概要/摘要 人工耳蜗植入（CI）用户通常能够在安静的环境中保持对话。 然而，当多个人同时说话时，例如在一个大家庭晚餐或在一家餐馆， CI用户很难参与对话，并且经常退出或避免这种情况。 理想地，用于从信号中去除背景讲话者（“多路重合”）的CI算法将允许改善的通信性能。 理解和会话参与。虽然CI包含降噪（NR）算法， 当背景是多路重合时，这些算法是无效的。从前景讲话者中分离出串音 提出了两个重大挑战。首先，信号和噪声的频谱特性非常相似， 两者都是言语。第二，多说话者串音的频谱和时间特性随时间变化， 因此很难预测。 尽管存在这些挑战，我们还是开发了一种非常有效的算法，称为SEDA， 胡言乱语SEDA提高了对在所有信噪比（SNR）下的多路串音中的语音的理解， 平均为26个百分点（或38个百分点，当相对于安静的听力标准化时）。相比之下， 商业NR算法（来自AdvancedBionics的ClearVoice）提供很少或没有可检测的益处。 在成功的第二阶段，我们产生了一个商业上可行的SEDA实现。然而，尽管如此， 要使SEDA进入商业化阶段，还需要做大量的工作。以下目标是在 与CI制造商合作，促进CI制造商获得SEDA技术许可。 目标1：在非串音收听情况下评估SEDA。至少，SEDA必须是有益的 具有多路重合，并且如果它要商业地实现到CI中，则在其他收听情况下是无害的。 因此，我们将使用语音识别来评估SEDA在非多路重合听觉场景中的效果， 收听者偏好和计算度量。 目标2：询问SEDA相对于CI制造商的商业产品的好处。我们 我将比较SEDA与Advanced Bionics、MED-EL、Cobriar和Oticon的NR的有效性 医学上关于在牙牙学语、白色和语音形状噪声中低估语音。 目标3：从SEDA的家庭评估中获得真实世界的反馈。我们会送病人回家 与SEDA合作一个月，以收集反馈，并确定意外问题或听力情况， 处理。 目标4：量化计算权衡对SEDA性能的影响。我们将修改 SEDA中用于调整计算要求的参数数量。使用计算度量 和语音识别，我们将评估这些变化对SEDA性能的影响。

PROJECT SUMMARY / ABSTRACT Cochlear implant (CI) users are typically able to maintain conversations in quiet environments. However, when multiple people are talking simultaneously, such as at a large family dinner or in a restaurant, CI users have great difficulty participating in conversations and frequently withdraw or avoid the situation. Ideally, CI algorithms to remove background talkers (“babble”) from the signal will allow for improved comprehension and conversational engagement. Although CIs incorporate noise reduction (NR) algorithms, these algorithms are not effective when the background is babble. Separating babble from a foreground talker poses two significant challenges. First, the spectral properties of the signal and noise are extremely similar as both are speech. Second, the spectral and temporal properties of multi-talker babble change with time and are therefore difficult to predict. Despite these challenges, we developed an extremely effective algorithm called SEDA to remove babble. SEDA improved understanding of speech in babble at all signal-to-noise ratios (SNRs) tested by an average of 26 percentage points (or 38 points, when normalized with respect to hearing in quiet). In contrast, a commercial NR algorithm (ClearVoice from Advanced Bionics) provided little to no detectable benefit. In a successful Phase 2, we produced a commercially viable implementation of SEDA. Nevertheless, significant work is required to bring SEDA to commercial readiness. The Aims below were developed in conjunction with CI manufacturers to facilitate SEDA technology for licensing by CI manufacturers. Aim 1: Evaluate SEDA in non-babble listening situations. At minimum, SEDA must be beneficial with babble and not detrimental in other listening situations if it is to be commercially implemented into a CI. Therefore, we will evaluate the effect of SEDA in non-babble auditory scenes using speech recognition, listener preference, and a computational metric. Aim 2: Interrogate benefits of SEDA relative to commercial offerings from CI manufacturers. We will compare the effectiveness of SEDA with NR from Advanced Bionics, MED-EL, Cochlear, and Oticon Medical on understating speech in babble, white, and speech-shaped noise. Aim 3: Obtain real-world feedback from at home evaluations of SEDA. We will send patients home for a month with SEDA to collect feedback and to ascertain unexpected issues or listening situations to be addressed. Aim 4: Quantify the effects of computational trade-offs on SEDA performance. We will modify the number of parameters used in SEDA to adjust the computational requirements. Using a computational metric and speech recognition, we will evaluate the effects of the of these changes on SEDA’s performance.