Speaker identification with machine learning algorithms

With the continuous development of technology, the number of automatic systems and devices entering our lives is increasing day by day. Different control mechanisms are needed for the security, control, dispatch and management of systems and devices controlled by these new technologies. In this search, methods that help distinguish characteristic features such as fingerprint recognition, retina recognition, face recognition, and speech recognition, which vary from person to person, have gained importance. In this study, a data set was created with speech samples taken from 10 different speakers aged between 18 and 39. Each speaker was recorded using equipment to say 25 different words one by one. A final recording was created by saying these 25 words consecutively. Thus, 260 different data consisting of 26 different records per person were obtained. Values of each data based on time, frequency and other characteristics were extracted with the Matlab program. The relationship of each of these extracted features with the desired user was examined with Pearson correlation analysis. As a result of Pearson correlation analysis, the features most associated with the output were determined and a ranking was made. Afterwards, tables were created using all 14 features, the first five and the first two, according to this order. Then, wavelet transform analysis was performed for the same data set using the wavelet transform interface of the Matlab program. Machine learning algorithms have been trained with all the data obtained so far and subjected to a process of making a prediction about the speaker. At the end of this process, the percentage success performances of different machine learning algorithms, as a result of their training with sound feature extraction and wavelet transform techniques, were analyzed and interpreted. According to the results obtained, as the number of features decreases, the accuracy of the predictions also decreases. In the data obtained as a result of the analysis made with wavelet transform, it was revealed that the entropy value can also be considered as a feature number.