Diagonal Factorization Subspace in I-Vector Extraction for Fast Computation and Memory Efficiency: A Case Study on Frog Sound Identification

In this study, we investigate the performance of low-dimensional identity vectors, called I-vectors, in automatically identifying frog species on the basis of their bio-acoustics. I-vector is one of the state-of-the-art techniques and has greatly improved speaker recognition performance. However, the I-vector method has some limitations, such as slow computation and high memory consumption. To address these issues, this study improves the I-vector algorithm using diagonal techniques. Although the proposed method reduces computational cost, it leads to some accuracy degradation. Therefore, an enhanced version, called Diagonal Factorisation Subspace in I-vector (DFS I-vector), is developed. Additionally, channel compensation fusion is implemented to further improve the accuracy of the I-vector approach. Frog species identification are employed as a case study for this research and experiment data for this study are employed from five databases, i.e. Frog Identification Expert System Database, Frogs of Australia Database, Frog Watch Database, British Library Amphibian Database and Intelligent Biometric Group Universiti Sains Malaysia Database, which consist of 2,656 samples of bio-acoustic syllables. Experimental results prove the promising output of the proposed DFS I-vector-32, achieving 90.67% accuracy, which is five times faster, and reducing memory consumption by one-fifth the memory of standard I-vector-16. Experimental results also demonstrate the advantage of the fusion approach, DFS I-vector-32 with WCCN(LDA), achieving 91.56% accuracy. Both performances are better than that of standard I-vector 16, which achieves 88.00% accuracy.