Designing better hearing aids using physiologically inspired speech enhancement

助听器可以深刻地改变听力障碍者的生活，但仅在英国就有大约600万这样的人没有使用助听器。一个重要的原因是，传统的助听器不能很容易地将语音等有意义的信号与背景噪音区分开来。功能齐全的耳朵和大脑的结合非常擅长这项工作，在即使只有少量噪音的情况下，它的表现也轻松超过了最好的计算机化语音识别应用程序。传统的助听器不加区别地放大语音和噪音，因此，即使大脑的神经通路可能没有受到损害，区分语音和噪音的任务也变得困难得多。人们已经尝试了各种自动语音增强的方法，但没有一种方法能接近自然的能力。尽管人们需要更好的解决方案，但最近的研究和开发进展缓慢，还没有出现突破性的技术。语音增强策略通常是基于数学原理而不是基于生理原理来开发的。这些方法虽然基于完全不同的策略，但有两个共同点：第一，它们主要依赖于信号的局部能量来操作信号特征，第二，它们并没有提高语音的清晰度。在这里，我们建议通过开发基于生理学原理的噪声中语音问题的工程解决方案来克服这一概念障碍。同样的技术也将有益于自动语音识别系统，因为两者的问题相似。我们预计在未来5年内，我们的工作将直接应用于助听器。在这一领域，世界上最大的两家公司(西门子助听器和谷歌信号处理)表现出了对这一方法的兴趣、支持和信心。我们在整个开发周期中拥有丰富的经验：我们发明、设计、评估和实施了一项降噪方案，该方案将成为下一代人工耳蜗有限公司人工耳蜗的一部分。我们在这项提议中的中心假设是，大脑在区分有意义的信号和噪音时使用稀疏编码，并使用动态词典来表示声音。我们将在听觉脑干的单个神经元中研究这种编码机制，并将根据结果开发新的信号处理策略。我们预计这些算法将比传统算法更好，因此可以帮助听力受损的人。动物模型对这个项目至关重要，因为不可能研究人类听觉脑干单个神经元的反应。稀疏神经元适应它们的反应，因为它们有一个有限的动态速率，它们不断地优化以响应环境，以减少冗余和最大化信息流。在这个项目中，我们将扩展对静态神经反应模式的描述，包括时变和上下文敏感成分，我们将测量脑干单个神经元的这些动态反应。了解神经元在噪声中的反应如何变化，将使我们能够创建可用于稀疏的动态词典。我们期望在这种表示中，语音和噪声是可分离的。我们将使用这些动态响应模式作为一种新的变换和稀疏的基础，以增强与理解相关的语音成分，从而在不降低质量的情况下提高语音的清晰度。我们将在大量的临床试验中对该算法进行评估。

Hearing aids can profoundly transform the lives of people with hearing impairments but in the UK alone about 6 million such people do not use them. An important reason for this is that conventional hearing aids don't make it easy to distinguish meaningful signals such as speech from background noise. The combination of a fully functioning ear and the brain are fantastically good at this job and easily outperform the best computerized speech-recognition apps when even just a small amount of noise is present. A conventional hearing aid amplifies both the speech and noise indiscriminately, so even though the neural pathways of the brain may be unimpaired the task of distinguishing speech from noise becomes much harder. Various approaches to automatic speech enhancement have been tried but none comes close to what nature can do.Despite the demand for better solutions, recent progress in research and development has been slow and no breakthrough technology has yet emerged. Speech enhancement strategies have been generally developed on the basis of mathematical, but not physiological principles. These methods, although based on fundamentally different strategies, have two things in common: first, they operate on signal features that rely primarily on the signals' local energy, and second, they have not improved speech intelligibility. Here, we propose overcoming this conceptual barrier by developing engineering solutions to the speech-in-noise problem that are based on physiological principles. The same technology will also be of benefit for automatic speech recognition systems, since the problems of both are similar.We expect to see direct applications of our work to be implemented in hearing aids within the next 5 years. Two of the biggest companies in the world in their field (Siemens for hearing aids and Google for signal processing) demonstrate the interest, support and the confidence toward this approach. We have substantial experience with the whole development cycle: we have invented, designed, evaluated and implemented a noise reduction scheme that will be part of the next generation of Cochlea Ltd. cochlear implants. Our central hypothesis in this proposal is that the brain uses sparse coding when distinguishing meaningful signals from noise and it uses a dynamic dictionary for sound representation. We are going to investigate this coding mechanism in individual neurons in the auditory brainstem, and based on the results, will develop novel signal-processing strategies. We expect that these algorithms will be better than conventional algorithms and consequently can help hearing impaired. An animal model is essential to this project, because it is impossible to study responses of individual neurons from the auditory brainstem in humans. Sparse Neuron adapt their response because they have a limited dynamic rate which they constantly optimize in response to the environment in order to reduce redundancy and to maximise the information flow. In this project, we are going to extend the description of static neural response patterns to include a time varying and context sensitive components and we will measure these dynamic responses in single neurons in the brain stem. Knowing how neuronal responses change in noise will enable us to create a dynamic dictionary that can be used for sparsification. We expect that in this representation speech and noise is separable. We will use these dynamic response pattern as the basis of a novel transformation and sparsification in order to enhance the components of speech that are relevant for understanding, thus improving speech intelligibility without reducing the quality. We will evaluate the algorithm in substantial clinical trials.