论文部分内容阅读
从语音信号预测伴随头动时,基于隐Markov模型(hidden Markov model,HMM)的头动合成方法的效果依赖于头动模式的划分和头动模式的正确识别。该文尝试了不同头动模式划分方法的头动合成效果。由于语音和头动之间是非确定性的多对多的映射关系,很难用固定的类别描述清楚,因此该类方法的头动模式识别率不高,头动合成效果受限。该文尝试采用逆传播(back-propagation,BP)神经网络的非线性回归方法,通过学习语音与头动之间的映射关系,实现语音信号到头动参数之间的直接连续映射,避免了HMM方法中头动模式不明确、头动模式识别错误带来的负面影响。实验表明,基于BP神经网络的回归方法有效地提高了语音到头动预测的准确度和头动合成的自然度。
From the perspective of speech signal prediction, the effect of head motion synthesis based on hidden Markov model (HMM) relies on the classification of head motion mode and the correct recognition of head motion mode. This paper attempts to analyze the head-and-tail combination of different head-and-head modes. Due to the nondeterministic many-to-many mapping relationship between speech and head movements, it is hard to describe clearly in fixed categories. Therefore, the head recognition rate of this kind of method is not high, and the effect of head movement synthesis is limited. This paper attempts to use the non-linear regression method of back-propagation (BP) neural network to learn the direct and continuous mapping between speech signals and head motion parameters by learning the mapping relationship between speech and head motion, avoiding the problem of HMM method In the first move mode is not clear, the first move the wrong pattern to identify the negative impact. Experiments show that the regression method based on BP neural network can effectively improve the accuracy of speech to head motion prediction and the naturalness of head motion synthesis.