In this paper, we report the results of automatic speech recognition experiments which used a model in which we regarded the recognition of observed phoneme as recall from association among the feature parameters of preceding, observed and following phonemes, and in which the neural network model was trained to learn correlations among their feature parameters. In these experiments, we adopted symmetrically connected three-vowel sets, Vi VqVj, as an example of continuous speech, with the central vowel, Fo, as the object to be recognized. Based on these experiments, it was shown that the proposed model, which utilized the contribution from the preceding and following phonemes, is useful for automatic continuous speech recognition.