Measurement of feedback in voice control and application in predicting and reducing stuttering using machine learning

语音控制中的反馈测量以及使用机器学习预测和减少口吃的应用

• 批准号: 2223533
• 金额: --
• 依托单位: University College London
• 依托单位国家: 英国
• 项目类别: Studentship
• 财政年份: 2019
• 资助国家: 英国
• 起止时间: 2019 至 无数据
• 项目状态: 已结题
• 来源: https://gtr.ukri.org/projects?ref=studentship-2223533
• 关键词: Measurement; feedback; voice; control; application

The PhD investigates: (1) how the brain encodes feedback from the speaker's own speech; (2) the biomarkers that predict moments of pathological stuttering; and (3) a neurophysiological monitoring system for identifying these features to trigger neuro-stimulation as part of a clinical/assistive technology. The work is at the intersection between engineering and psychological research. The PhD will use digital signal processing (DSP) and machine learning (ML), to explore applications of assistive and healthcare technologies to motor-speech disorders.I work with neuro-stimulation and electroencephalography (EEG) equipment. I use DSP to extract spectral features of EEG signals which indicate the neural oscillations of the underlying network. As part of my BSc in Psychology, I used transcranial alternating current stimulation. This entrains neural networks to oscillate a specific rate. The aims was to yield a frequency specific modulation to speech. My past work has furthered my understanding in image and signal processing and experimental neuroscience.The PhD is specifically interested with how perturbations to the feedback system alter the brain's response. Combined EEG+fNIRS provides rich data about brain activity. Through EEG+fNIRS during speech studies, the PhD aims to gain insight into the neural mechanisms during alterations to recurrent speech feedback. Using novel ML techniques allows the extraction of biomarkers of fluency to be determined. Support vector machines (SVM) and neural networks (NN) will be used to extract features that predict disfluency. These learning methods are suitable as they are resistant to outliers relative to logistic regression. Using Least absolute shrinkage and election operator (LASSO) with SVMs and multi-class NN's (such as a deep convolutional NN) would allow the evolution of features across dimensions (i.e. Electro-physiological and hemodynamic). Also, these biomarkers can guide design of assistive and clinical technologies. It is unlikely that the cortical and hemodynamic responses of individuals would create a valid set of super-features. Instead, the ML techniques allows idiosyncratic features (aka biomarkers) appropriate for fluency states for individuals. The incorporation of engineering techniques will not only allow a further understanding of the underlying biology but also how to address associated pathological states. Speech requires the integration of different systems including memory, motor co-ordination and linguistic planning, amongst other factors. Therefore, The current PhD work will require of image and signal processing, artificial intelligence, computational neuroscience, technology and their integration across disciplines. Namely the integration of DSP and AI to trigger assistive technology and/or non-invasive brain stimulation with the intention of reconfiguring the associated neural networks to restore fluency. The work has implications about how feedback systems are used in fluent speech production; combines machine learning, psychology, human computer interaction and neuroscience and provides background on applications of AI to clinical approaches to neurological disorders.

博士研究：（1）大脑如何编码来自说话者自己的语音的反馈;（2）预测病理性口吃时刻的生物标志物;以及（3）用于识别这些特征以触发神经刺激的神经生理监测系统，作为临床/辅助技术的一部分。这项工作是在工程和心理学研究之间的交叉点。博士将使用数字信号处理（DSP）和机器学习（ML），探索辅助和医疗保健技术在运动语言障碍中的应用。我使用神经刺激和脑电图（EEG）设备。我使用DSP来提取脑电信号的频谱特征，这些特征表明底层网络的神经振荡。作为我心理学学士学位的一部分，我使用了经颅交流电刺激。这会使神经网络以特定的速率振荡。目的是产生对语音的频率特定调制。我过去的工作加深了我对图像和信号处理以及实验神经科学的理解。博士特别感兴趣的是反馈系统的扰动如何改变大脑的反应。EEG+fNIRS组合提供了关于大脑活动的丰富数据。通过在语音研究期间的EEG+fNIRS，博士旨在深入了解在改变经常性语音反馈期间的神经机制。使用新的ML技术可以提取流畅性的生物标志物。支持向量机（SVM）和神经网络（NN）将被用来提取预测不流利的特征。这些学习方法是合适的，因为它们相对于逻辑回归抵抗离群值。将最小绝对收缩和选举算子（LASSO）与SVM和多类NN（例如深度卷积NN）一起使用将允许特征跨维度（即，电生理和血液动力学）的演变。此外，这些生物标志物可以指导辅助和临床技术的设计。个体的皮质和血液动力学反应不太可能产生一组有效的超级特征。相反，ML技术允许适合个人流畅状态的特质特征（又名生物标志物）。工程技术的结合不仅可以进一步了解潜在的生物学，还可以解决相关的病理状态。言语需要不同系统的整合，包括记忆、运动协调和语言规划等因素。因此，目前的博士工作将需要图像和信号处理，人工智能，计算神经科学，技术及其跨学科的整合。即DSP和AI的集成，以触发辅助技术和/或非侵入性脑刺激，目的是重新配置相关的神经网络以恢复流畅性。这项工作对如何将反馈系统用于流利的语音产生了影响;结合了机器学习，心理学，人机交互和神经科学，并提供了人工智能应用于神经系统疾病临床方法的背景。