摘要
提出了一种基于蚁群聚类神经网络的汉语耳语音声调识别方法.根据耳语音发音特点,以听神经平均发放率、幅值包络、共振峰、声道长度等构成的多维矢量描述声调特征,采用蚁群聚类算法将多维特征矢量聚类后,送入局部有监督特征映射神经网络进行声调识别.这一方法通过对特征参数的聚类压缩了神经网络的输入神经元数目,因而可以有效避免在大数据条件下神经网络不易收敛及速度慢的问题.对多人耳语音声调的识别实验显示,采用蚁群聚类神经网络的耳语音声调识别方法与传统方法相比,性能明显提高,平均正识率达到87.5%.
Based on analysis of acoustic and perception characteristics of whispered mandarin speech, a tone detection method using ant colony clustering is proposed. A multi-dimension feature vector consisted of amplitude envelope, formant, vocal tract length, average firing rate of auditory nerves is chosen as the mainly cue for whispered tone. The feature vectors are clustered by ant colony algorithm and then input to regional supervised feature mapping neural network for training and recognizing. The experiment results show that 87.5% average recognition accuracy could be reached and the performance of proposed method is improved significantly compared with classical models.
出处
《应用科学学报》
CAS
CSCD
北大核心
2008年第5期511-515,共5页
Journal of Applied Sciences
基金
国家自然科学基金(No.60572076)
江苏省高校自然科学基金(No.05KJB510113)资助项目
关键词
耳语音
声调检测
蚁群聚类
神经网络
whispered speech, tone detection, ant colony clustering, neural network