Age-group and gender identification from speech in Azerbaijani language

Abstract

Our Speech comprises of paralinguistic features such as: identity, age, gender, accent etc. The objective of this paper is to identify, the age-group and gender of the speaker from Azerbaijani speech. The paper will be focused on both the adult and children speech identification. The identification of the age and gender of children speech is more complex than the adult’s speech as the voice of boys and girls before puberty coincide and additionally the puberty complicates the distinction between an adult and a teenager which results in possible errors with age-group identification. Moreover, the existence of numerous accents of Azerbaijani data an additional milestone. To identify age and gender from the speech the data should be pre-processed and then the features extracted. Next step is to classify according to results obtained. Various approaches are going to be tested such as x-vectors and i-vectors that are based on Deep Neural Network architecture. Then there is MFCC - a feature extraction technique a part of Automatic voice processing for unique feature extraction. On top of that the GMM-SVM model which is a Gaussian mixture model is run. KNN and MLP are another prominent approach to be used as a classifier for age and gender identification problem. Another feature extraction technique called SDC – shifted delta cepstral coefficient will be tested and compared with the MFCC results. The music and audio analysis package called Librosa and a PyAudio library are used to enable the record and play of an audio for demonstration purposes in the future. There outcome of the model is going to be classified into 4 age groups which are: Children (7-14), Young aged (15-24), Middle aged (25-54) and Seniors (55-80) and 2 genders: Male and Female. A proper identification of age and gender is sometimes a hard task for a human being as well which complicated the identification process for the machines.