Human-machine interaction is a popular field of study and its importance is increasing day by day. Detecting emotions from speeches has also been an exciting topic in the field of humanmachine interaction. With the spread of technologies such as digital personal assistants, speech-totext models and call center systems, which are the applications of speech, the systems that can detect emotions from speeches that can be seen as the next natural step of this process have become more valuable. In addition, since speech is a physiological process and it is reflected in the speeches of people's mental and physical health, emotion detection studies can be used to evaluate physical, and mental health status. In this study, an approach that uses the features of cochleagram and convolutional neural networks (CNN) was proposed to detect emotions from speech. In the proposed approach, speech recordings are converted to overlapping cochleagrams. It is then converted to a grayscale image. Colored images are created with obtained grayscale images through color maps. After each image is set to the same size, it is classified with the CNN network. The proposed approach provided a 2.8% improvement in the classification performance of the data set, which corresponds to a relative error reduction rate of 19.46% compared to state-of-the-art techniques.