基于深度学习模型的语音特征提取方法研究

With the rapid development of Mobile Internet, speech recognition which remains key to human-machine interaction is attracting more and more attention. Especially in the era of big data when access to large amount of speech data is possible, exploring ways to manipulate these unlabeled data effectively has become the hot topic in the field of speech recognition. Meanwhile, deep learning models find more opportunities in combination with speech recognition technologies because of its outstanding performance in unlabeled speech corpus processing and data modeling. This thesis is based on the combination of speech recognition and deep learning theory, aims at extracting more robust acoustic features with application of deep models. In the experiment, both Auto-Encoder and Deep Neural Network are constructed to extract new features from mfcc, respectively through techniques of unsupervised and supervised feature learning. Finally, the new features obtained from deep models mentioned above have increased the word recognition accuracy by 1.96% and 3.53%, compared with mfcc features.