Prediction of speaker characteristics with hybrid spectral features from human voice

This study focuses on predicting the personal characteristics of speakers using voice data. It has great potential in a wide range of applications, from forensic cases to automated voice response systems. The main purpose of the research; Speaker recognition, gender and age group estimation based on the voice recordings of 24 volunteers. In the analyses performed on the sound data, Mel-frequency cepstral coefficients (MFCC) were removed as time/frequency space hybrid features, while basic frequencies and formants were determined as frequency space features. These attributes are then combined into an attribute pool. The obtained features were used using four different machine learning algorithms to generate support vector machines (SVM), k-nearest neighbors (KNN), gradient boosting and classification, and regression trees; It was evaluated for age group and gender estimation. The results show that the predictions of age groups obtained by the support vector machines algorithm can be performed with 93% accuracy, and the gender predictions can be performed with 99% accuracy. In the speaker recognition task, the SVM algorithm achieved high success with 93% accuracy. This high accuracy reveals that the attributes extracted from the audio data can provide reliable information about the personal characteristics of the speaker, especially using the SVM algorithm. These findings point to potential applications in forensic processes and biomedical applications, for example in determining the age and sex of a speaker. Such techniques offer important areas of research and application that can be used in many fields, from automated voice answering systems to social analysis. Keywords: Gender Estimation, Age Group Estimation, Support Vector Machine, K-Nearest Neighbors, Gradient Boosting, Classification and Regression Tree