【摘要】 语音识别领域中所采用的经典HMM模型,忽略了语音信号间的相关信息。针对这一问题,利用语音信号的空间相关性对经典HMM模型进行补偿,得到一种改进模型。该方法通过空间相关变换,描述了当前语音特征与历史数据之间的空间相关性,从而对联合状态输出分布进行建模。改进模型的解码算法利用空间相关性变换的参数更新算法在经典HMM的解码算法基础上得到。实验结果表明,上述方法在说话人无关连续语音识别系统上获得了明显的性能改进。更多还原

【Abstract】 The conventional HMM used in speech recognition ignores the correlation information in speech signal. In order to solve this problem,an improved acoustic model is obtained by compensating the conventional HMM with the spatial correlation information in speech signal. By using the spatial correlation transformation (SCT),the approach describes the spatial correlation between the current speech feature and the history data,in order to model the joint state output distribution. The decoding algorithm of the improved model is obtained based on the decoding algorithm of the conventional HMM and the parameter updating algorithm of SCT. Experimental results show that the above approach improves the performance of speaker independent recognition systems significantly.更多还原