Speech recognition adaptation for low resource populations

自动语音识别是移动的设备和消费电子产品的基本属性。不幸的是，由于这些系统是用成人语音训练的，因此在儿童和说话困难的人使用时表现不佳。缺乏来自这些群体的可用训练语音使得为他们开发模型变得困难。我们将研究有效的模型自适应方法，使用最少的训练数据来适应现有的成人语音识别模型，用于儿童和说话困难的人。预期的成果将改善澳大利亚人使用自动语音识别系统的机会，这些人目前与语音控制的环境和教育设备的通信受到限制。

Automatic speech recognition is an essential attribute of mobile devices and consumer electronics. Unfortunately, as these systems are trained with adult speech, they perform poorly when used by children and people with speaking difficulties. The lack of available training speech from these groups makes developing models for them difficult. We will investigate efficient model adaptation methods that use minimal training data to adapt existing adult speech recognition models for use with children and people with speaking difficulties. The intended outcomes will improve access to automatic speech recognition systems for Australians whose communication with speech-controlled environmental and educational devices is currently restricted.