| 摘要: |
| 将声学特征与韵律特征相结合,提出一种新的混合区间特征,并将该特征和常见的美尔倒谱系数(MFCC)特征与线性预测倒谱系数(LPCC)特征进行对比,通过符号化语言辨识方法对北方方言、吴方言、粤方言和闽方言进行辨识,以验证混合区间特征的有效性。结果表明,混合区间特征比MFCC特征和LPCC特征具有更好的方言辨识效果,对4种汉语方言15s语音片段的方言辨识率可以达到92%。4种方言中,混合区间特征对闽方言和粤方言的识别率最高,分别达到了96%和95%。 |
| 关键词: 语音辨识 汉语方言 韵律特征 声学特征 GMM符号化器 |
| DOI： |
| 投稿时间：2007-03-28修订日期：2007-07-05 |
| 基金项目:江苏省“十五”社科基金项目(K3-013);江苏省高校自然科学基金项目(99KJB510002)资助 |
|
| A New Features of Chinese Dialects Indentification |
|
GU Ming-liang1,2
|
| (1.School of Physics and Electronic Engineering, Xuzhou Normal University, Xuzhou, Jiangsu 221116, China;2.Linguistic Institution, Xuzhou Normal University, Xuzhou, Jiangsu, 221116, China) |
| Abstract: |
| Combining acoustic features with prosodic features, this paper presents a new hybrid block feature.In order to test the efficiency of the new feature, comparative experiments are done on the speech database consisting of North, WU, YUE and MIN dialects.The experimental results show that the new feature can performs better than traditional MFCC and LPCC features.An average accuracy of 92% is achieved in four Chinese dialects with 15 seconds speech segments.And the identification accuracy of MIN and YUE dialects is best in four dialects.They are 96% and 95% respectively. |
| Key words: identification Chinese dialects prosodic features acoustic feature GMM tokenizer |
