This paper reports the building of a limited vocabulary speech recognition system for Sora without a speech database designed specifically for automatic speech recognition (ASR). The system was built using speech data collected in field for acoustic phonetic analysis of the Sora language in Assam, India. As Sora is an under resourced language, the speech database is small and contains recordings of single words. Thus, the system is trained without a language model. There is no available ASR for Sora language and hence, this is the first attempt to build one for the language which may lead to the development of a more robust ASR for the language. The ASR shows better recognition rate with Subspace Gaussian Mixture Model (SGMM) and Deep Neural Network (DNN) frameworks. While the system performs with adequate accuracy, as expected, phonetically similar words are often misrecognized.