Traditional speech recognition systems have relied on power spectral densities, Mel-frequency cepstral, linear prediction coding and formant analysis. This paper introduces two novel input feature sets and their extraction methods for intelligent phoneme identification. These input sets are based on intrinsic phonetic characteristics of Arabic speech comprising of the dimensionally reduced Power Spectral Densities (DPSD) and Location, Trend, Gradient (LTG) values of the captured speech signal spectrum. These characteristics have been subsequently utilized as inputs to four different neural network based recognition classifiers. The classifiers have been tested for twenty-eight Arabic phonemes utterances from over one hundred non-native speakers. The results obtained using the proposed feature sets have been compared and it has been observed that LTG based input feature set provides an average phoneme identification accuracy of 86% as compared to 70% obtained through applying DPSD based inputs for similar classifiers. It is worthwhile to note that the methods proposed in this paper are generic and are equally applicable to other regional languages such as Persian and Urdu.