SPAM e-mails have a direct cost in terms of time, server storage space, network bandwidth consumptions and indirect costs to protect privacy and security breaches. Efforts have been done to create new filters techniques to block SPAM, however spammers have developed tactics to avoid these filters. A constant update to these techniques is required. This paper proposes a novel approach which is a characters-word-based technique. This approach uses a multi-neural networks classifier. Each neural network is trained based on a normalized weight obtained from the ASCII value of the word characters. Results of the experiment show high false positive and low true negative percentages.