The spoken language understanding is an emerging field between speech processing and language understanding from the human spoken utterance. It can apply in different sectors for representing the meaning of a human spoken utterance in different domains. The analysis and identification of the speech act and intent in the use of language for the speaker motivation in political speech candidates are important factors for the willingness of the speaker and the audience. But there is no an annotated speech act and intent recognition corpus for the Amharic language from the spoken utterance, due to these, the politician transmits never known what is say and what is meant in the transmission and the audience also not creates a common understanding of the transmitted information, because of the different interpretations of the utterance. In this study to tackle the problems, we have been proposed a convolutional neural network approach with a word and sentence embedding technique for analyzing and identifying a once utterance speech act and intention in political speech candidate. We have been used a word2vec approach for converting each word in the corpus for vector values with a continuous bag of word architecture and also used a mean embedding approach for converting each utterance value in numeric value representation. We have applied a convolutional neural network approach, for extracting deeply features from the distributional matrix representation of each speech utterance and also for the identification of each speech act and intention utterance class by using the different activation functions. We have achieved an accuracy of 92.5%, 89.3%, and 83.8% for speech act, speech intent and the intent based speech act classification as Softsign function of SoftMax classifier, and also 88.1% and 56.8% for speech act and the intent based speech act classification as ReLu function and 59.6% for speech intent as Softplus function for sigmoid function in the output layer respectively.