用于语音识别的空间相关性变换

Spatial correlation transformation for speech recognition

针对经典隐含Markov模型忽略了语音信号之间的依存关系的问题,提出一种线性特征变换——空间相关性变换,利用同一个说话人的不同语音单元之间的相关性(空间相关性)得到鉴别性能更好的新特征。该变换的最优变换矩阵在最小协方差准则下得到。识别系统采用新特征及其模型参数代替原特征及其模型参数进行Viterbi搜索。实现空间相关性变换的关键是最优变换矩阵的计算,提出了两种相应的算法。实验结果表明:该方法在说话人无关识别系统上取得了比自适应方法更好的性能,同时该方法与自适应方法结合应用可进一步提高系统性能。

The traditional Hidden Markov model for speech recognition ignores the relationships between speech signals. This paper presents a linear feature transformation, Spatial Correlation Transformation, to utilize the correlation between different acoustic units of the same speaker （Spatial Correlation） to obtain new features having better discrimination. The optimum transformation matrix is determined based on the Minimum Covariance criterion. The recognition system uses these new features and the corresponding model parameters in the Viterhi search instead of the original features. The key to the transformation is the calculation of the optimum transformation matrix. Experiments show that this approach achieves better performance than adaptation approaches on the speaker independent recognition system. Moreover, the combination of this approach and adaptation approaches further improves the system performance.