<p>The classification and determination of a natural language derived from a specific content and dataset goes through a process called Spoken Language Identification (LID). LID normally entails the procedure of data processing whereby valuable features are extracted; this is a mature process in which the standard LID features have previously been developed utilising the MFCC (Mel-Frequency Cepstral Coefficients), SDC (Shifted-Delta-Cepstra), GMM (Gaussian Mixture Model), and a framework based on i-vector. Nevertheless, improvement (optimisation) is still needed in the aspect of process of learning so as to enable complete capturing of all the extracted features’ embedded knowledge. ELM (Extreme Learning Machine) is considered as one of the most powerful machine learning algorithms for performing regression and classification and is highly effective in training a single hidden layer neural network. However, the ELM’s learning process is still not fully optimised due to the selection of random weights embedded in the input hidden layer. In the context of this study, grounded upon the standard feature extraction, ELM is chosen as the learning model for LID. The ESA-ELM (Enhanced Self-Adjusting-ELM) is one of the model’s optimisation approaches used as the benchmark; rather than adopting EATLBO (Enhanced Ameliorated Teaching Learning-Based Optimisation), the model utilises an improved kidney-inspired algorithm. Ultimately, this enhanced version of the ESA-ELM is referred to as Improved Kidney-Inspired Algorithm-ELM (IKA-ELM). The generation of the outcomes was carried out based on two different LID datasets (Dataset 1 and Dataset 2) where Dataset 1 contains eight languages while Dataset 2 contains ten languages. The outcomes indicated that the IKA-ELM outperformed the ESA-ELM, ELM, SVM, and RF classifiers on both datasets with an accuracy of up to 97.50% on Dataset 1 and 97.20% on Dataset 2.</p>