Increasing trust & reliability of AI sign language translation

全世界有4.66亿人有听力损失。我们提出的解决方案通过提供英国手语（BSL）的实时翻译，有助于打破沟通障碍。我们的项目是一个使用基于视觉的机器学习（ML）来翻译车贴语的移动应用程序。我们的创新在于使用这项技术并创造性地训练一个强大的机器学习模型，以支持它改善听力障碍人士的可及性。迄今为止，机器学习系统的有限可靠性并没有阻碍手语应用程序的可访问解决方案的发展。由于车贴语是一种复杂的3D语言，因此缺乏ML应用程序的鲁棒性——手势、身体姿势、面部表情和上下文都会影响符号的选择，其中个人偏见（口音）和地区方言同样适用。这进一步加剧了数据语料库不足，无法为通用的BSL翻译器训练机器学习。缺乏准确可靠的机器学习系统也会导致用户缺乏信任和遵从性。作为以人为本的设计师，我们感到这种沮丧，并欣赏边缘群体无法获得的直观和可访问的用户体验。这一领域缺乏技术发展意味着目前的解决方案提供商诉诸于亲自使用人工口译员或通过视频会议提供按需服务。这是非常昂贵的，限制了听障人士的自由，并最终限制了他们的生活质量。我们的应用程序有潜力用于广泛的用例，但是为了分级和限制技术挑战，我们首先关注咖啡店的订购。例如，它可以用于餐馆、零售、教育和任何视频会议软件，帮助人们与有听力障碍的人交流并理解他们。我们已经准备好在2023年6月1日之前开始，并计划在6个月内为一个用例提供一个可访问的、以人为中心的ML应用程序，并计划使用训练数据（包括我们自己的和开源的语料库）来提供理想的体验。当公司采用该技术并更普遍地发展业务时，这种验证和公开将使我们能够将该技术发展到其他用例。简而言之，目前的挑战是，目前可用的工具无法可靠地翻译手语。在这个项目中，我们将构建有效训练ML模型的工具。

466 million worldwide have hearing loss. Our proposed solution helps to break down communication barriers by providing a real-time translation of British Sign Language (BSL). Our project is a mobile application that uses vision-based machine learning (ML) to translate BSL. Our innovation lies in the use of this technology and creatively training a robust ML model that underpins it to improve accessibility for people with hearing-impairments.To date, the limited reliability of ML systems has not stunted the development of accessible solutions for sign language applications. Since BSL is a complex, 3D language the robustness of ML applications is lacking - hand gestures, body posture, facial expression and the context all influence the choice of signs where individual bias (accent) and regional dialect similarly apply. This is further compounded by an insufficient corpus of data to train the ML for a general-purpose BSL translator.The lack of accurate and reliable ML systems also leads to a lack of trust and compliance from users. As human-centred designers, we feel this frustration and appreciate intuitive and accessible user experiences that are not available for marginalised groups.The lack of technical development in this area means current solution providers resort to using human interpreters in person or as an on-demand service over video conferencing. This is very costly and restricts the freedom of hearing-impaired individuals and ultimately, their quality of life.Our application has the potential to be used for a wide range of use cases, but to stage-gate and constrain the technical challenge, we first focus on ordering in coffee shops. For example, it could be used in restaurants, retail, education, and any video conferencing software to help people communicate with and understand those with hearing-impairments.We are fully ready to begin before the 1st June 2023 and over 6 months plan to deliver an accessible, human-centric, ML application for one use case and a plan for to use the training data - both our own and open-source corpora - to deliver a desirable experience. This validation and exposure will enable us to grow the technology to other use cases as companies adopt the technology and grow the business more generally.In short, the current challenge is that the tools available at the moment aren't capable of reliably translating sign language. In this project, we are going to build the tools to effectively train an ML model.