A novel technique of oriented local features (OLF) for speech recognition has been introduced in this paper. A speech recognition system for dysphonic patients is implemented by application of the proposed technique. The developed system is evaluated by performing the training in three different ways: (a) with pathological samples, (b) with normal samples, and (c) with pathological and normal samples together. We compare the performance of the proposed feature with the most widely used speech feature in speech recognition, i.e., Mel-frequency cepstral coefficients. The Hidden Markov model is used for recognizing the speech. The proposed technique achieved a 94.98 recognition rate, which is almost identical to the recognition rate of 95.45 obtained with MFCC. © 2015 ISCA.