State of the art speech recognition systems (ASR), mostly use Mel-Frequency cepstral coefficients (MFCC), as acoustic features. In this paper, we propose a new discriminative analysis of acoustic features, based on spectrogram analysis. Both spectral and temporal variations of speech signal are considered. This has improved the recognition performance especially in case of noisy situation and phonemes with time domain modulations such as stops. In this method, the 2D Discrete Cosine Transform (DCT) is applied on small overlapped 2D Hamming windowed patches of spectrogram of Spanish phonemes and enhanced by means of bi-cubic interpolation. An adaptive strategy is proposed for the size of patches over the time to construct unique length vectors for different phonemes. These vectors are classified based on K-nearest neighbor (KNN) and linear discriminative analysis (LDA) and reduced rank LDA (RLDA). Experimental results demonstrate improvement in recognition performance for noisy speech signals and stops.