Self-supervised learning (SSL) representations derived from multilingual speech using transformers have significantly improved the performance of automatic speech recognition (ASR) tasks in low-resource languages. This study focuses on the development of an ASR system for the Malayalam language using pretrained SSL representations and enhance its performance with the inclusion of a statistical n-gram language model. A language model enhances speech recognition by improving contextual accuracy by guiding the decoding process with linguistic probabilities and reducing transcription errors. Our findings indicate that the language model augmentation reduces the word error rate by 12.5% on in-domain test data and 5.9% on out-of-domain test data. The datasets used are diverse, encompassing a variety of speakers, speech domains, and recording environments. Our results underscore the critical role of augmenting pre-trained speech representations with language models in developing ASR systems for languages with limited speech resources. The model is made openly available on HuggingFace for reproducibility and adaptation to similar languages.