Improved voice control and transcription for wearable electronic devices

大多数人认为，智能手表等可穿戴电子设备的用户体验因语音-机器界面不佳而受挫。由于语音识别易受风噪声和重叠说话者的影响，这是由设备无法基于其位置区分声音引起的，用户需要将设备移动到非常靠近他/她的嘴。在许多情况下，此操作会阻止在语音控制期间查看屏幕。我们将研究使用萨里大学开发的盲源分离（BSS）技术改进语音控制和转录的可行性。我们工作的目标是与单麦克风系统相比，将未经训练的单词识别错误率降低50%。这项工作是高度创新的，因为传统的源分离技术需要胶囊分离太大（> 15厘米）的智能手表应用。成功将把口头输入命令和文本的便利性转变为日常设置中的可穿戴设备。

By most accounts the user experience with wearable electronic devices such as smartwatches is frustrated by the poor voice-machine interface. Due to the vulnerability of speech recognition to wind noise and overlapping talkers, caused by the devices' inability to discriminate sounds based on their location, the user is required to move the device very close to his/her mouth. In many situations this action prevents the screen from being viewed during voice control. We will investigate the feasibility of improving voice control and transcription using a blind source separation (BSS) technology developed at the University of Surrey. The objective of our work will be to reduce the untrained word recognition error rate by 50% compared to a single-mic system. This work is highly innovative because conventional source separation techniques require capsule separations that are too large (>15cm) for smartwatch applications. Success will transform the ease of inputting commands and text verbally into wearable devices in everyday settings.