基于韵律信息的语种辨识系统

提出利用超声层次上的韵律信息来进行语种辨识.在子词分割的基础上结合元音/辅音检测模型,将语音分割为CⁿV的假音节模型,然后对每个假音节提取一个五维的特征韵律参数.分割结果比传统的分割方法更接近实际的音节结构.在韵律系统中,对所有的语言建立高斯混合模型(GMM),研究模型的离散度和模型间的干扰程度.实验表明,韵律模型对不同的语言、不同的语系和同一语系中不同语言均具有一定的区分效果.

Prosodic Information Used in Language Identification

This paper deals with automatic language identification based on prosodic features and rhythmic information.A segmentation method has been used to divide the speech signal into sub-word units,and with vowel/consonantal detection,segment the speech into vowel units and consonantal units.Pseudo-syllable units are then produced.Several parameters are extracted and modeled from pseudo-syllable and modeled with a Gaussian mixture model (GMM).Experiments are performed on free and telephone channel speech.The results show that the sub-word system and the rhythmic system are effective.