A new method for the automated speech production assessment (ASPA) of hearing impaired children is presented in this paper, providing feedback about the pronunciation quality of words and sentences uttered during unsupervised practice in the course of speech development. A database of the sounds produced by hearing impaired subjects was set up and assessed with a subjective test. The Mean Opinion Score (MOS) obtained in this way constituted the reference for automated assessment. The essence of the ASPA method is the joint assessment of sound and rhythm errors. After several methods were tested, the output activity of the neural networks trained to classify speech sounds was used to assess sound correctness. Dynamic time warping, adapted to the speech of the hearing impaired, was used to determine rhythm errors. ASPA provides input data for an expert system for the selection of the next word to be practiced. The novelty of the procedure is that it provides a method for the assessment of non-typifiable pronunciation errors. Results were compared with individual expert assessment and subjective tests. Automated assessment surpassed the overwhelming majority of subjective assessors and approximated the correctness of individual expert assessment. Our ASPA method is implemented in our “Speech Assistant” application, which also provides a language-independent sound visualization module and is successfully applied to assist the hearing impaired.

中文翻译：

---

听力障碍儿童的自动言语产生评估

本文提出了一种听力障碍儿童自动语音生成评估 (ASPA) 的新方法，提供有关语音发展过程中无监督练习中说出的单词和句子的发音质量的反馈。建立了听力受损受试者产生的声音数据库，并通过主观测试进行评估。以这种方式获得的平均意见得分（MOS）构成了自动评估的参考。ASPA 方法的本质是对声音和节奏错误的联合评估。在测试了几种方法之后，训练用于对语音进行分类的神经网络的输出活动用于评估声音的正确性。动态时间扭曲适用于听力受损者的语音，用于确定节奏错误。ASPA 为专家系统提供输入数据，用于选择要练习的下一个单词。该程序的新颖之处在于它提供了一种评估非典型发音错误的方法。结果与个别专家评估和主观测试进行了比较。自动化评估超越了绝大多数主观评估者，接近个别专家评估的正确性。我们的 ASPA 方法是在我们的“语音助手”应用程序中实现的，该应用程序还提供了一个独立于语言的声音可视化模块，并成功应用于帮助听力障碍者。结果与个别专家评估和主观测试进行了比较。自动化评估超越了绝大多数主观评估者，接近个别专家评估的正确性。我们的 ASPA 方法是在我们的“语音助手”应用程序中实现的，该应用程序还提供了一个独立于语言的声音可视化模块，并成功应用于帮助听力障碍者。结果与个别专家评估和主观测试进行了比较。自动化评估超越了绝大多数主观评估者，接近个别专家评估的正确性。我们的 ASPA 方法是在我们的“语音助手”应用程序中实现的，该应用程序还提供了一个独立于语言的声音可视化模块，并成功应用于帮助听力障碍者。