Acknowledging printed characters remains a significant challenge within the field of optical character recognition (OCR), especially for complex scripts and multi-lingual documents. Arbitrary initialization weights and biases applied at input in Extreme Learning Machine(ELM) can trigger suboptimal performance and poor classification accuracy. To correct it, this study proposes an optimized ELM framework using three advanced metaheuristic algorithms: Swarm intelligence (PSO), grey wolf-inspired search (GWO), and the parameter-free JAYA algorithm. These models are evaluated on a dataset of printed odia character images, where intra-class similarity and inter-class variability pose challenges. Aforesaid limitations are subdued through experimental results that demonstrate the optimized ELM variants which outperform the baseline ELM in precision, training efficiency, and stability. This work exhibits intelligent systems for OCR applications and exploring optimization-enhanced neural classifiers in multilingual character recognition tasks. The outcome is demonstrated by ELM-GWO with accuracy of 98.62 while 98.02 and 97.31 for ELM-JAYA and ELM-PSO respectively for research in multilingual character recognition.