Machine Learning for Hearing Aids: Intelligent Processing and Fitting

目前的助听器有两大局限：1)助听器音频处理策略不灵活，不能充分适应听力环境；2)听力测试和助听器试配程序不能对听力损失的潜在性质进行可靠诊断，并经常导致设备配戴不佳。本研究计划将使用新的机器学习方法来彻底改变助听器技术的这两个方面，导致智能助听器和测试程序能够主动了解患者的听力损失，从而实现更个性化的试配。智能音频处理助听器的最佳音频处理策略取决于声学环境。例如，在安静的办公室里进行的对话，应该以不同的方式处理，而不是在熙熙攘攘的餐厅里进行。当前的高端助听器确实基于监视传入音频的简单方面的简单声环境分类系统在少量不同的处理策略之间切换。然而，分类精度有限，这是听力设备在噪声多源环境中表现非常差的原因之一。未来的智能设备应该能够识别更大、更多样化的音频环境，可能是通过智能手机进行无线通信。此外，助听器应该使用这些信息来告知助听器中声音的处理方式。该项目的第一个分支的目的是开发促进这类设备开发的算法。其中一个重点将是一类被称为音频纹理的声音，这种声音结构丰富，但信号在时间上是均匀的。例子包括：用餐者在餐馆里喋喋不休；火车在铁轨上嘎嘎作响；风在树林里咆哮；水龙头里流水。音频纹理通常表示环境，因此它们携带了有关场景的有价值的信息，可以通过助听器加以利用。此外，纹理经常破坏目标信号，它们的抑制可以帮助听力受损的人。我们将开发高效的纹理识别系统，可以识别环境中存在的噪音。然后，我们将设计和测试定制的实时降噪策略，利用环境中存在的音频纹理信息。智能听力设备感觉神经性听力损失可能与许多潜在原因有关。在耳蜗内，可能存在内毛细胞(IHC)或外毛细胞(OHC)功能障碍、代谢障碍和结构异常。理想情况下，听力专家应该根据对听力损失的根本原因的详细了解来安装患者的助听器，因为这决定了最佳的设备设置或是否继续进行干预。不幸的是，目前配对程序中使用的听力测试，称为听力图，不能可靠地区分许多不同形式的听力损失。需要更复杂的听力测试，但事实证明，很难设计这些测试。在该项目的第二个环节中，我们提出了一种不同的方法，即在测试的每个阶段之后改进患者听力损失的模型，并使用该模型来自动设计和选择下一阶段特别有用的刺激。这些测试将是快速、准确的，并能够确定患者特定的潜在功能障碍的形式。然后，患者听力损失的模型将被用来以最佳方式安装听力装置，使用计算机模拟和听力测试的混合。

Current hearing aids suffer from two major limitations:1) hearing aid audio processing strategies are inflexible and do not adapt sufficiently to the listening environment,2) hearing tests and hearing aid fitting procedures do not allow reliable diagnosis of the underlying nature of the hearing loss and frequently lead to poor fitting of devices.This research programme will use new machine learning methods to revolutionise both of these aspects of hearing aid technology, leading to intelligent hearing devices and testing procedures which actively learn about a patient's hearing loss enabling more personalised fitting. Intelligent audio processingThe optimal audio processing strategy for a hearing aid depends on the acoustic environment. A conversation held in a quiet office, for example, should be processed in a different way from one held in a busy reverberant restaurant. Current high-end hearing aids do switch between a small number of different processing strategies based upon a simple acoustic environment classification system that monitors simple aspects of the incoming audio. However, the classification accuracy is limited, which is one of the reasons why hearing devices perform very poorly in noisy multi-source environments. Future intelligent devices should be able to recognise a far larger and more diverse set of audio environments, possibly using wireless communication with a smart phone. Moreover, the hearing aid should use this information to inform the way the sound is processed in the hearing aid. The purpose of the first arm of the project is to develop algorithms that will facilitate the development of such devices.One of the focuses will be on a class of sounds called audio textures, which are richly structured, but temporally homogeneous signals. Examples include: diners babbling at a restaurant; a train rattling along a track; wind howling through the trees; water running from a tap. Audio textures are often indicative of the environment and they therefore carry valuable information about the scene that could be harnessed by a hearing aid. Moreover, textures often corrupt target signals and their suppression can help the hearing impaired. We will develop efficient texture recognition systems that can identify the noises present in an environment. Then we will design and test bespoke real-time noise reduction strategies that utilise information about the audio textures present in the environment.Intelligent hearing devicesSensorineural hearing loss can be associated with many underlying causes. Within the cochlea there may be dysfunction of the inner hair cells (IHCs) or outer hair cells (OHCs), metabolic disturbance, and structural abnormalities. Ideally, audiologists should fit a patient's hearing aid based on detailed knowledge of the underlying cause of the hearing loss, since this determines the optimal device settings or whether to proceed with the intervention at. Unfortunately, the hearing test employed in current fitting procedures, called the audiogram, is not able to reliably distinguish between many different forms of hearing loss. More sophisticated hearing tests are needed, but it has proven hard to design them. In the second arm of the project we propose a different approach that refines a model of the patient's hearing loss after each stage of the test and uses this to automatically design and select stimuli for the next stage that are particularly informative. These tests will be be fast, accurate and capable of determining the form of the patient's specific underlying dysfunction. The model of a patient's hearing loss will then be used to setup hearing devices in an optimal way, using a mixture of computer simulation and listening test.