深度自动编码器在耳语音说话人特征提取中的应用
首发时间:2015-04-07
摘要:针对耳语音没有基频且能量低、易受噪声,导致特征提取困难的问题,提出一种基于深度自动编码器(Deep Autoencoder,DAE)的耳语音特征提取方法。DAE是一种无监督的生成模型,通过多层抽象可以直接从数据中学习到特征。学习到的特征由于经过多层抽象,是一种高层的特征,相比低层特征更稳定且具有融簇性;并且DAE具有很好的噪声鲁棒性,可以减少噪声的影响。实验表明:在传统的MFCC-GMM识别方法下,耳语音说话人的识别效果稍差,并且在发音方式不匹配的测试中,识别率大幅下滑;而采用深度自动编码器作为特征提取器可以得到更好的识别效果。
For information in English, please click here
An Application of Deep Autoencoder in Feature Extraction of Speaker Recogniton with Whispered Speech
Abstract:Due to the profound differences between whisper and neutral speech,the performance of speaker identification systems of MFCC-GMM method trained with neutral speech degrades significantly.The solution of extracting a better feature is Deep Autoencoder,a generative model with multi-layers. The results show that the performance of the seamless neutral/whisper mismatched speaker recognition system based on MFCC-GMM is far behind expectation while the recognition rate rises with DAE method.
Keywords: deep learning whispered speech MFCC feature extraction
论文图表:
引用
No.4636705101310314****
同行评议
共计0人参与
勘误表
深度自动编码器在耳语音说话人特征提取中的应用
评论
全部评论0/1000