Abstract:

The bottleneck features extracted from deep neural network not only have long term contextdependence and compact representation of speech signal, but also can replace the traditional MFCC features for GMM-HMM acoustic modeling. The authors apply bottleneck features and their concatenated features with MFCC into Lhasa-Tibetan continuous speech recognition. The experiments in Lhasa-Tibetan continuous speech recognition show that the concatenated features of bottleneck features and MFCC achieve better performance than the posterior features of deep neural network and mono-bottleneck features.

Key words: Lhasa-Tibetan, continuous speech recognition, GMM-HMM, bottleneck features, deep neural network (DNN)

摘要：

基于从深度神经网络提取的瓶颈特征具有语音长时相关性和紧凑表示的特点, 将瓶颈特征及其与MFCC的复合特征用于藏语连续语音识别任务中, 可以代替传统的MFCC特征进行GMM-HMM声学建模。在藏语拉萨话连续语音识别任务中的实验表明, 瓶颈特征的复合特征取得比深度神经网络后验特征和单瓶颈特征更好的识别表现。

关键词: 藏语拉萨话, 连续语音识别, 高斯混合–隐马尔科夫模型, 瓶颈特征, 深度神经网络