An Efficient Speaker Identification System with High-Level Feature Extraction and Database Dimensionality Reduction.

Speaker identification is a biometric technology that leverages distinct characteristics obtained from vocal utterances to verify users' identities. Later advancements in multiple fields have raised the importance of speaker identification systems, particularly in security applications. The challenging task in speaker identification systems is how accurately to extract discriminative features from the speech signal. This paper presents a novel approach method that integrates the two-dimensional discrete multi-wavelet analysis-based critical sampling scheme (2D-DMWT-CS) with the principal component analysis (PCA) to employ a reliable and efficient speaker identification system. The proposed method incorporates four phases: preprocessing, feature extraction, dimensionality reduction, and training and classification. During the preprocessing phase, successive refinement techniques such as duration division, silence removal, resampling, and dimension reshaping are applied to the databases. All databases speech samples are then analyzed using the 2D-DMWT-CS. The resultant discriminative features of the wavelet analysis are further processed by the PCA during the supplementary dimensionality reduction phase. The latter provides high-level, hierarchically ordered features that come with a substantial benefit for enhancing the classification accuracy of the convolutional neural network (CNN). The suggested approach was validated by testing and evaluating the framework over many individuals using their speech identities in four online datasets: RAVDESS, TIMIT, ELSDSR and SALU-AC. The achievement results, in terms of the recognition rate, were 97.19% for the TIMIT database, 97.96% for the RAVDESS database, and 98.91% for the ELSDR database, which are higher results than those in the state-of-the-art literature. The reliable and efficient identification rates with high accuracy and fast learning with reducing dimensionality, are the main contributions of this work. [ABSTRACT FROM AUTHOR]