Subvocal Speech for Augmentative and Alternative Communication

 描述（由申请人提供）：由于需要为无法通过发声进行交流的人提供更有效的辅助和替代交流（AAC）设备，因此提出了本阶段II SBIR。该项目遵循我们的初步工作，令人信服地证明，从语音清晰度肌肉记录的表面肌电图（sEMG）信号可以提供一种新的有效的交流形式，而无需发声。由于基于表面肌电信号的语音识别不依赖于声道的声学激励，因此它很容易适用于识别亚声（即，口部）语音。因此，对于喉切除术的患者来说，无声的言语是一种明显的替代交流形式。该项目的目标是提供一个在Android移动终端（智能手机）上运行的预商用、可穿戴、无声语音识别（SSR）系统，该系统可以为喉切除术后的非说话者提供在家中、社区或通过电话进行免提、可理解的通信的能力。该项目处于直接二期开发的良好位置。已经在两个方面实现了原理证明和降低的风险：i）无线传感器设计已经成功地在基本原型中实现，其改进了记录来自面部和颈部的8个关节肌肉的sEMG信号的任务;和ii）迄今为止，最先进的SSR引擎已经被制定为能够从2000个单词的词汇表中准确识别出无声的连续语音，这些词汇表在未受损的人身上进行了测试扬声器以及2人喉切除术。第二阶段将通过将必要的传感器设置减少到面部肌肉部位来推进这些技术，这些传感器将被集成到一个预商用设备中，供喉切除术的非扬声器使用。Aim 1将单个sEMG传感器整合到一个舒适的面部界面中，并将所采集的信号联合收割机合并到一个数据流中，以便通过蓝牙连接到运行SSR软件的Android设备。由此产生的数据采集系统将被封装，实验室测试，并与喉切除术的受试者进行评价。Aim 2将为喉切除术用户创建一个先进的SSR引擎，将面部传感器的必要数量从8个减少到4个子集，同时以低于10%的错误率实现1000个单词的识别性能。该创新的影响在于，它为喉切除术使用者提供了一种替代形式的语音，该替代形式a）克服了依赖于麦克风的当前自动语音识别（ASR）系统的限制，B）与需要手持接触的电子喉技术相比是免提的，c）不像当前的语音假体那样具有差的可懂度或对手术干预和维护的需要，以及d）易于用作AAC设备的人机接口。

 DESCRIPTION (provided by applicant): This Phase II SBIR is prompted by the need for more effective Augmentative and Alternative Communication (AAC) devices for persons unable to communicate through vocalization. The project follows our preliminary work, which convincingly demonstrated that surface electromyographic (sEMG) signals recorded from speech articulation muscles can provide a new and effective form of communication without vocalization. Because sEMG-based speech recognition does not rely on acoustic excitation of the vocal tract, it is readily applicable to recognizing subvocal (i.e. mouthed) speech. Subvocal speech is therefore an obvious alternative form of communication for patients with laryngectomy. The goal of this project is to deliver a pre-commercial, wearable, subvocal speech recognition (SSR) system operating on an Android mobile device (Smartphone) that can provide non-speakers with a laryngectomy the ability to produce hands-free, intelligible communication in the home, community, or over the phone. The project is well positioned for direct Phase II development. Proof-of- principal and reduced-risk have been achieved on two fronts: i) wireless sensor designs have been successfully implemented in a rudimentary prototype that improves the task of recording sEMG signals from 8 articulatory muscles of the face and neck; and ii) the most advanced SSR engine to date has been formulated to achieve accurate recognition of subvocal continuous speech from a 2000 word vocabulary tested on unimpaired speakers as well as from 2 people with laryngectomy. Phase II will advance these technologies by reducing the requisite sensor set to just facial muscle sites, which will be integrated into a pre-commercial device for use by non-speakers with a laryngectomy. Aim 1 will consolidate the individual sEMG sensors into a conformable facial interface and combine the acquired signals into a data stream for Bluetooth connectivity to the Android device running the SSR software. The resulting data acquisition system will be encapsulated, bench-tested, and evaluated on subjects with a laryngectomy. Aim 2 will create an advanced SSR engine for laryngectomy users that will reduce the requisite number of sensors from 8, to a sub-set of 4 on the face, while attaining a recognition performance for 1000 words at an error rate less than 10%. The impact of this innovation is that it provides laryngectomy users with an alternative form of speech that a) overcomes the limitations of current automated speech recognition (ASR) systems that are microphone dependent, b) is hands-free compared to electrolarynx technologies requiring handheld contact, c) does not suffer from poor intelligibility or the need for surgical interventio and maintenance as with current voice prostheses, and d) is readily adaptable as a man-machine interface for AAC devices.