腭裂语音高鼻音等级自动识别关键技术研究

Hypernasality is one the most typical characteristics of Cleft Palate (CP) speech. The grades of hypernasality reflect different open port sizes of velopharyngeal. It is an important indicator to evaluate operation effectiveness. Moreover, ti provides assistant diagnosis for the necessity of carrying out the follow-up operation and treatment. Currently, the evaluation of CP speech is mostly carried out by experienced speech therapists. However, it strongly depends on their clinical experience and subjective judgment. Automatic evaluation of CP speech provides an objective and assistant assessment for both CP patients and doctors. It has various clinical applications. Currently, the major bottleneck in the field of CP speech research is the collection of CP speech data, which is tremendously limited by the number of patients, the accent of speakers, and thoughtful design of vocabulary list, which highly reflects typical CP speech characteristics. The majority of current CP speech researches are based on small size of CP speech databases, which only include several vowels or words. In this work, a more extensive CP database is applied. The speech data are collected by the Department of Cleft Lip and Palate (DCLP), Hospital of Stomatology, Sichuan University, which has the largest number of CLP patients in China. The majority of current researches on CP speech hypernasality analysis detect the existence of hypernasality only. However, CP speech therapists are more interested in finding out different levels of hypernasality, which indicates the velopharyngeal gap size. Whereas, very few works have been done in this field. In this work, we propose an automatic CP speech hypernasality grades evaluation algorithm. We investigate the acoustic features extraction methods and pattern recognition algorithms to classify four hypernasality grades: normal, mild,moderate and severe. This is the first domestic research on this topic.

高鼻音是腭裂语音最常见的临床表现之一。对腭裂高鼻音的语音评估具有十分重要的临床意义，目前国际各唇腭裂治疗单位均通过专业语音师的主观判听实现，该方式受语音师经验及主观状态等影响因素较多。自动腭裂高鼻音等级识别将为腭裂语音评估提供客观辅助诊断，社会应用前景广泛。目前，国内外对腭裂语音信号处理研究仍属起步阶段，其最主要瓶颈为腭裂语音数据采集。课题针对该瓶颈，与国内最大最权威的腭裂语音治疗中心合作，建立样本广泛的权威腭裂语音数据库。目前，国内外对腭裂高鼻音语音研究停滞在对高鼻音有无的判别上，临床意义不大。临床实践中，医生与语音师更希望得到对高鼻音等级的识别。课题首次实现普通话腭裂语音高鼻音四个等级（正常、轻、中、重度）自动识别的关键问题研究：从发音机理探索腭裂高鼻音声学特征的形成；研究腭裂高鼻音等级敏感声学特征参数；探讨腭裂高鼻音声学特性与模式识别算法的结合。课题开展弥补了国内在该领域的研究空白。

高鼻音是腭裂语音障碍中最常见的表现之一。在临床上，高鼻音的评估主要通过语音师的主观判听实现。主观评估主要依赖于语音师的临床经验，耗时耗力。自动的高鼻音等级评估能为临床上腭裂治疗提供辅助诊断。目前国内外腭裂语音处理的主要瓶颈为腭裂语音数据库的采集。同时，腭裂高鼻音语音研究停滞在对高鼻音有无的判别上，临床意义不大。本课题旨在研究腭裂高鼻音四类等级（正常、轻度、中度和重度）的自动判别算法。本课题经过5年的研究，收集了丰富的腭裂语音数据库。基于腭裂语音发声过程来建模，寻求对高鼻音等级敏感的特征参数。本课题提取了四大类别的特征参数：基于声门信号的特征、基于声道特性的特征、基于共振峰的特征、以及基于整个发音过程的特征。同时结合浅层分类器和深度学习技术实现高鼻音等级的自动判别。除了高鼻音以外，我们也扩展研究了其它的腭裂语音障碍，包括鼻漏气、喉塞音、辅音省略和咽擦音的自动判别。本课题旨在为临床提供高效、高准确率的辅助诊疗，满足腭裂临床医生的需求，具有无创、实时、客观、高效、识别准确率高的特点。我们的研究有效弥补了国内在腭裂语音自动评估算法的空白。