论文部分内容阅读
本文描述了在语音识别中特征提取的一种方法。此种方法采用1/3倍频程实时频谱分析仪给出的随时间变化的频谱——三维语音频谱作为原始特征参数,对三维语音频谱分别进行频率域一阶差分,频率域二阶差分,时间域一阶差分和拉普拉斯变换,并分别进行二值化处理得到各种不同的特征参数。然后对这些特征参数进行突出过渡频谱的非线性时间域规正,得到多种特征矩阵。最后用实验方法评定了各种特征矩阵对语音识别的贡献,挑选了两种特征矩阵作为船舶驾驶台语音识别实验系统的特征矩阵。
This article describes a method of feature extraction in speech recognition. In this method, the frequency-domain three-dimensional speech frequency spectrum given by the 1/3 octave real-time spectrum analyzer is taken as the original characteristic parameter, and the three-dimensional speech frequency spectrum is subjected to the first-order frequency domain difference and the second- First-order difference and Laplace transform in time domain, and binarized respectively to obtain various different characteristic parameters. Then, the non-linear time domain regularization of the prominent transition spectrum of these characteristic parameters is performed, and a variety of characteristic matrices are obtained. Finally, the contribution of various feature matrices to speech recognition was evaluated experimentally, and two feature matrices were chosen as the feature matrices of the experimental system of ship’s bridge speech recognition.