<p>Recently, there has been a significant trend of moving from a deep neural network-based hybrid approach to an end-to-end (E2E) approach for automatic speech recognition (ASR). While E2E models achieve state-of-the-art results in most benchmarks for high-resource languages, it is always difficult to train decent E2E models with low-resource languages. Given the recent success of self-supervised learning (SSL), which does not require labeled data to pretrain a representation for downstream tasks, we present an ASR system for the Manipuri language in the news domain via multilingual speech representations in this paper. We follow the state-of-the-art wav2vec2 XLSR approach which uses multilingual speech representations learned by a pretrained model to train a fine-tuned model. The pretrained model was fine-tuned with different configurations and amounts of data, with their effects on inference provided. Furthermore, an n-gram language model is used to improve the performance of the system with the limited size of the dataset. Additionally, recognition error analysis is discussed. We observe the best word error rate (WER) and character error rate (CER) on test sets obtained by decoding with a language model as 8.25% and 2.43%, respectively.</p>