采用深层神经网络中间层特征的关键词识别

Keyword Spotting Based on Deep Neural Networks Bottleneck Feature

在基于模板匹配的关键词识别中,提出采用深层神经网络的中间层特征(bottleneck,BN)作为特征输入,将其取代传统的声学参数来生成后验概率图.首先采用传统语音识别的过程训练一个中间层很窄的深层神经网络,将所有的语音特征经过这个神经网络后得到稳健的BN特征;然后利用混合高斯模型将BN特征转化成后验概率图;在识别过程中,利用后验概率图作为特征参数,采用简化的分段动态时间规整算法实现关键词匹配.在TIMIT数据库上,相对于采用传统感知线性参数的系统,采用BN特征的系统,识别准确率有30%的提升.

In this paper, the BN （ bottleneck ） features extracted from DNN （ Deep Neural Networks ） are adopted to replace the tradi-tional acoustic features in template-based Keyword Spotting. Firstly a traditional speech recognition DNN with narrow bottleneck istrained,then the acoustic features are transformed to BN feature through this BN feature extractor. The BN features are fed into aGMM （ Gaussian Mixture Model ） to generate Gaussian posteriorgrams, which will be served as the input of segmental DTW （Dynam-ic Time Warping ）. A language independent keyword spotting experiments are carried in TIMIT corpus. Experimental results demon-strate that the BN features can outperform the conventional PLP （ Perceptual Linear Prediction ） acoustical features, with an absoluterecognition accuracy improvement of 30%.