In this paper, a method of synthesizing speech facial expression by the Kohonen neural networks (KNN) is proposed, and the result of quantitative investigation of synchronicity with the speech is reported. The proposed method generates facial expression series in speech of vowel to vowel with KNN, which uses the still images of the 5 vowels pronunciation and non-speech state. Synchronization with speech sound was done by separating the stable sections and transition sections of facial expression based on the sound power, assigning a fixed facial expression to the stable section and the evenly distributed facial expression series to the transition section. Verification test was conducted on two synchronization methods: one that assigns an evenly distributed facial expression series to the intervals between syllables, and the proposed method in this paper which assigns an evenly distributed facial expression only to the transition section. Comparison was based on the per-frame difference between the synthesized images and the actual speech image. The proposed method yielded more realistic speech facial expression than the other method.