Nepali is a language with various dialects, complex syllable structure, and rich morphology. It is spoken in Nepal, India, and Bhutan by wider range of population. Being a low-resource language, it possesses a challenge in the development of an accurate Automatic Speech Recognition system using deep learning techniques. Recently, Conformer-based architectures have been widely adopted for speech recognition due to their ability to exploit local dependencies while maintaining global interactions and context, which shows Conformer is the perfect architecture for modeling Nepali ASR systems. However, despite their effectiveness, the development of ASR models for Nepali face challenges such as limited training resources and suboptimal performance in handling the language’s unique phonetic and linguistic characteristics. To address these challenges, we propose NepConformer, an End-to-End ASR system for Nepali. The model is implemented and trained using NVIDIA’s NeMo framework on the SLR54 Nepali speech dataset, which consists of \(\approx \) 157 K transcribed speech samples. NepConformer achieves a new state-of-the-art Character Error Rate of 6.01% and a Word Error Rate of 23.96%. These results demonstrate the effectiveness of Conformer-based models in tackling low-resource ASR tasks and pave the way for more accurate and accessible speech recognition systems for Nepali speakers.