论文部分内容阅读
针对传统语音识别应用中识别效果不理想的问题,提出一种基于美尔谱系数(MFSC)特征的有限局部权重共享卷积神经网络(CNN)语音识别。首先,借鉴图像处理中对输入图像信息的处理方式,构建语音信号的二维阵列特征映射输入形式,每个映射表示为含静态数据、一阶导数、二阶导数的MFSC系数特征,便于应用图像处理方式进行识别;其次,引入图像处理的卷积神经网络,并且针对语音信号特征的局部特性,构建有限局部权重共享卷积神经网络学习算法,提高语音信号辨识度并降低算法复杂度;最后,通过实验对所提算法进行验证,并且给出算法参数变化影响实验,为具体应用提供依据。
Aiming at the problem that the recognition effect is not satisfactory in traditional speech recognition applications, this paper proposes a limited local weighted shared convolution neural network (CNN) speech recognition based on the MFSC feature. First of all, using the processing method of input image information in image processing, the two-dimensional array feature mapping input form of speech signal is constructed, each mapping is represented as MFSC coefficient with static data, first derivative and second derivative, Then, a convolution neural network based on image processing is introduced. According to the local characteristics of speech signal features, a limited local weight sharing convolution neural network learning algorithm is constructed to improve speech signal recognition and reduce the complexity of the algorithm. Finally, The proposed algorithm is validated experimentally, and experiments on influence of algorithm parameters variation are given to provide basis for specific application.