基于空间相关性变换的声学模型训练

为了在语音识别中增强对不同语音单元之间的相关性的利用，该文基于空间相关性变换(Spatial Correlation Transformation，SCT)框架，提出一种新的模型训练算法，在说话人无关模型的训练中利用训练数据中的空间相关性进行模型参数重估。该算法对所有训练数据进行空间相关性变换，削弱数据间的空间相关性，使重估的模型更不依赖训练数据，以改善模型的性能。实验表明，基于空间相关性变换框架的模型训练方法与基于该框架的特征变换方法相结合，使系统的平均错误率相对基线系统下降了18%。

Acoustic Model Training Based on Spatial Correlation Transformation

In order to enhance the utilization of the correlation between different acoustic units in speech recognition, a novel model training approach based on the Spatial Correlation Transformation (SCT) framework is proposed in this paper, in which the speaker-independent model parameters are re-estimated using the spatial correlation information in the training data. In this algorithm, SCT is applied to all training data, to decrease the correlation among the training data, make the model re-estimated less dependent on the training data, and then improve the performance of the model. Experiments show that the combination of SCT-based model training and SCT-based feature transformation achieves a relative reduction of 18% of average syllable error rate compared to the baseline system.