| 本文已被：浏览 4534次 下载 6131次 |
码上扫一扫！ |
|
|
| 频域动态特征提取中的多层信道正规化 |
|
王东1,2, 朱小燕1,2, 刘盈1,2
|
|
1.清华大学,智能技术与系统国家重点实验室,北京,100084;2.清华大学,计算机科学与技术系,北京,100084
|
|
| 摘要: |
| 语音识别领域已经取得了稳步发展并出现了众多实用系统,但众所周知,今天的识别技术还远没有达到要求,而"鲁棒性"问题是系统性能提高的一个主要障碍.集中讨论了一种对抗语音识别系统脆弱性的通行方法--信道正规化技术,提出了一种新的正规化策略--多层信道正规化MLCN(multi-layer channel normalization)新的算法应用递归补偿算法,在频谱域和倒谱域两层上进行正规化,降低噪音和去除信道畸变,从而为后续识别过程提供更鲁棒的特征参数.在此基础上,探讨了一种新的语音识别特征参数的提取-频域动态倒谱系数,由于MLCN的引入,频域的动态信息被恰当地集成到最终的特征向量中.在gallina系统中的实验证明了这种新方法的有效性. |
| 关键词: 语音识别 特征提取 Mel倒谱系数 信道正规化 频域动态特征 |
| DOI： |
| 分类号: |
| 基金项目:Supported by the National Natural Science Foundation of China under Grant No69982005 (国家自然科学基金); the NationalGrand Fundamental Research 973 Program of China under Grant No.G 199803050703 (国家重点基础研究发展规划项目(973)) |
|
| Multi-Layer Channel Normalization for Frequency-Dynamic Feature Extraction |
|
WANG Dong,ZHU Xiao-Yan,LIU Ying
|
| Abstract: |
| Despite the steady progress made in the area of speech recognition and a high number of practicalapplications, it is widely acknowledged that recognition technology today is not at the desired level. One mainobstacle is what said "robustness". This paper focus on one popular idea in antagonizing speech systemvulnerability-channel normalization, and presents a new normalization algorithm MLCN (multi-layer channelnormalization), which exploits the recursive compensation progress in two domains (spectral domain and cepstraldomain) to depress the noise and channel distortion, so that the more robust speech representation for the followingprocessing is achieved. A new frequency-dynamic feature extraction algorithm is also proposed due to theintroduction of MLCN, which allows dynamic information integrated in the final feature vectors. Experimentalresults of the gallina system demonstrate the validity of the new algorithm. |
| Key words: speech recognition feature extraction MFCC channel normalization frequency-dynamic feature |
