随着老龄化人口程度的加深，老龄化引起的听力损失问题也变得越加严重，影响了听力损失患者的日常的交流与安全。除了手术和药物治疗外，佩戴助听器是最有效的听力干预手段。目前，助听器技术正在面临着基于先进的数字信号处理、无线通信和人工智能技术带来的新的变革。为了解决复杂声学环境下用户体验不佳、直接使用手机进行网络音视频通话功能受限、电池续航能力不足和以及听力状况无法得到有效的跟踪调试等方面瓶颈，本文设计了一种基于智能终端的双耳助听声学处理系统，并开展了双耳助听下的相关算法研究；实现了基于智能移动的双耳助听功能。论文的具体贡献与创新点如下：1）提出了基于智能终端的双耳助听声学处理的系统方案。该系统移除了耳侧的DSP，并将音频信号处理转移到智能移动终端上，降低耳端功耗，延长了其电池的使用寿命。该系统从结构设计上实现了算法内置，为助听器的个性化处理提供便利，并使得软件算法维护和更新更加容易。2）提出了一种应用于双耳助听下的基于多通道的MMC混合语音增强算法。算法采用双重滤波进行信号的增强，前置滤波首先对目标信号做出估计，后置滤波通过耳间相关性处理来保留双耳线索。这种利用双重滤波处理的方式进行语音增强使得语音失真和噪声抑制之间取得折中，其泛化能力较好，可以在多种噪声场景下可以更有效的提高言语可懂度。3）提出了基于深度学习的语音可懂度增强算法。先后设计两种基于循环神经网络的算法框架，提取语音信号的多重维度上的特征，分别以不同的目标损失函数做训练处理，目标是在移动处理平台上对系统语音可懂度进行改善，并在多噪声维度下提高系统的鲁棒性。基于低复杂度的网络设计使得算法在计算资源和存储资源有限的移动平台上的应用成为可能；同时算法内引入两个级联的后置滤波器来重构语音信号，消除频带间噪声的同时保留其可听分量，使得最终输出的语音信号达到语音可懂度最优。4）搭建了基于智能终端的双耳助听声学处理平台，实现了前文所设计的基于深度学习的语音可懂度增强算法，在系统上对算法的处理性能进行了评估，得到了实时处理的结果和相关延时，证实了其增强效果的有效性。同时附加上纯音测听验配以获取听力损失程度的过程，完备了双耳助听声学系统的功能。
With the growing number of the aging population, the problem of hearing loss caused by aging has become more serious, affecting the daily communication and safety of hearing loss patients. In addition to surgery and medical treatment, wearing a hearing aid is the most effective hearing intervention. Currently, hearing aid technology is facing new challenges based on advanced digital signal processing, wireless communication, and artificial intelligence technology. To solve the bottlenecks of current hearing aids, this thesis proposed a binaural hearing acoustic processing system based on a smartphone and carried out research on multi-channel noise suppression algorithms and deep learning algorithm for speech enhancement; finally realized the proposed platform. The main contributions and innovations of this paper are as follows:1) A system-level solution of binaural hearing acoustic processing based on the smartphone is proposed. The system removes the DSP on the ear and transfers the audio signal processing to the mobile terminal, which reduces the power consumption of the ear and extends the battery life. The system has built-in algorithms from the structural design, which facilitates the personalized processing of hearing aids, and makes it easier to maintain and update software algorithms. The thesis analyzed the whole signal flow and measured the processing delay of each flow part separately, which is convenient for the optimization of each part of the system.2) A multi-channel hybrid speech enhancement algorithm for binaural hearing aids is proposed. The algorithm uses double filtering to enhance the signal. The pre-filtering first estimates the target signal, and the post-filtering preserves the binaural cues through inter-ear correlation, which makes a compromise between speech distortion and noise suppression. The proposed algorithm has the generalization ability, which can effectively improve speech intelligibility in various noise scenarios.3) A speech intelligibility enhancement algorithm based on deep learning is proposed. The thesis designed two kinds of algorithm frameworks based on recurrent neural network extracted the multi-dimensional features of the speech signal and trained them with different objective functions respectively. The design of the objective function is based on the metrics of speech enhancement, which belongs to the multi-objective learning criterion method. The goal is to improve the speech intelligibility of the system on a mobile processing platform and to improve the robustness of the system in multiple noise dimensions. The network design based on low complexity makes it possible to apply the algorithm to mobile platforms with limited computing and storage resources. At the same time, two cascaded post-filters are introduced into the algorithm to reconstruct the speech signal, which eliminates the noise between the bands and retain their audible components.4) A system-level solution of binaural hearing acoustic processing based on the smartphone is implemented, and the speech enhancement algorithm based on deep learning designed above is realized. The processing performance of the algorithm is evaluated on the system, and real-time processing is obtained. The results and related delays confirm the effectiveness of its enhancement effect. At the same time, the process of pure tone audiometry to obtain the degree of hearing loss is included, and the sub-channel loudness compensation is completed on the system according to the audiogram, which completes the function of the binaural hearing acoustic system.
