关键词: 梯度方向直方图, 语谱图映射, 说话人规整, 动态规划

Abstract: Spectrogram is an image reflecting time-varying spectral magnitude.The correspondence between spectrograms is established based on Gradient Orientation Histogram（GOH）to find the corresponding frequency structures，which benefits speaker normalization and further speech processing.Before extraction of feature parameters，the local feature in a spectrogram is described and the non-linear correspondence on the frequency axes between spectrograms of two speakers is established.In fact，the method is to find the optimal match by using dynamic programming given the similarity measure of two frequency bins.The experiments on the TIDIGITS corpus show reduction on the error rate under mismatched condition of training and testing data.

Key words: Gradient Orientation Histogram（GOH）, spectrogram correspondence, speaker normalization, dynamic programming