Fuzzy-Based Ensemble Feature Selection for Automated Estimation of Speaker Height and Age Using Vocal Characteristics

Abstract

Estimating speaker attributes from vocal characteristics is an important research area with applications in forensic science, biometric identification, and human-computer interaction. The accurate estimation of these attributes requires the effective extraction of relevant audio features from the audio signal. This work proposes a new approach for automatic speaker height and age estimation using fuzzy-based ensemble feature selection with speech parameters. This approach derives the initial feature importance from different feature selection (FS) methods. The obtained feature importance values are then sorted into a matrix, which is converted to ranks and passed to the fuzzy c-means (FCM) algorithm to produce the final feature ranking and identify the most distinctive features for estimation. The proposed approach can combine the results of multiple feature importance methods into an ensemble approach or choose the best features based on the feature importance from a single method without requiring a predefined number of top features. Several experiments were performed to evaluate the proposed approach on acoustic features obtained using the OpenSMILE toolkit from the TIMIT dataset. The results show that the proposed approach can effectively select the most informative features, and it outperforms similar studies on the same dataset, with promising results of 5.4, 4.71 mean absolute error (MAE) in height estimation and 5.38, 5.24 MAE for age estimation for males and females, respectively.