Speaker Age Classification and Regression Using i-Vectors

Abstract

In this paper, we examine the use of i-vectors both for age regression as well as for age classification. Although i-vectors have been previously used for age regression task, we extend this approach by applying fusion of i-vectors and acoustic features regression to estimate the speaker age. By our fusion we obtain a relative improvement of 12.6% comparing to solely ivector system.We also use i-vectors for age classification, which to our knowledge is the first attempt to do so. Our best results reach unweighted accuracy 62.9%, which is a relative improvement of 16.7% comparing to the best results obtained in age classification task at Age Sub-Challenge at Interspeech 2010. Index Terms: speaker age recognition, regression, classification, computational paralinguistics Acoustic features for age estimationFeature sets were extracted with openSMILE [14] and consist of 450 features for each utterance in aGender corpus [15]. Those acoustic, prosodic and voice quality features are low-level descriptors (e. g. MFCCs, LSP Frequency, F0, voicing probability,

Machine-Learning Analysis of Voice Samples Recorded through Smartphones: The Combined Effect of Ageing and Gender

“…Indeed, by comparing females included in the YA and OA groups as well as males included in the YA and OA groups, in separate analyses, we have examined the pure effect of ageing on voice. Our findings fully agree with previous reports demonstrating the effect of ageing on the human voice [ 24 , 25 , 26 , 27 , 28 , 33 , 34 , 35 , 36 , 37 , 38 ]. Early studies based on the qualitative/perceptual evaluation of voice recordings have demonstrated that physiologic ageing leads to several changes in specific characteristics of the human voice [ 1 ].…”

“…In our study, by applying the ROC curve analysis, we demonstrated in detail the high accuracy of our machine learning analysis in demonstrating age-related changes in the human voice. Our results fit in well with previous studies applying automatic classifiers based on machine learning analysis [ 24 , 25 , 26 , 27 , 28 , 33 , 34 , 35 , 36 , 37 , 38 ]. More in detail, our machine learning algorithm has achieved higher results than those obtained on the INTERSPEECH 2010 age and gender sub-challenge feature set [ 33 , 34 ].…”

“…In our study, SVM achieved relatively high performance with an accuracy of 95.3% in age recognition and of 95.5% in gender recognition, showing comparable or even better results than those obtained in previous reports [ 33 , 34 , 35 , 38 , 73 ]. When comparing our methodological approach to those previously used, it is important to consider that we started with a large dataset of features (more than 6000), adopting dedicated ranking and feature selection algorithms [ 33 , 34 , 35 , 36 , 37 , 38 , 73 ]. The advantages of applying those algorithms consist of obtaining smaller dataset of features (only 20 features in our study), easier math handled and with shorter computation time.…”

“…More in detail, machine learning can be applied to predict outcomes from recurring patterns of features within various types of multidimensional data sets [ 32 ]. Several authors have applied automatic classifiers based on machine learning analysis on voice recordings to classify healthy subjects according to their age and gender [ 24 , 25 , 26 , 27 , 28 , 33 , 34 , 35 , 36 , 37 , 38 ]. More recently, to further improve the overall accuracy of the machine learning analysis, several studies have included an increasing number of voice features in the datasets [ 24 , 25 , 26 , 27 , 28 ] and compared the performance of different machine learning algorithms [ 37 , 38 ].…”

Neural Regression Trees

“…To properly represent the speech signals, we adopt one of the most effective and well-studied representations, the i-vectors [30]. I-vectors are statistical low-dimensional representations over the distributions of spectral features, and are commonly used in state-of-the-art speaker recognition systems [31] and age estimation systems [32], [33]. Respectively, 400-dimensional and 600-dimensional i-vectors are extracted for Fisher and SRE datasets using the state-of-the-art speaker identification system [34].…”