The goal of the proposed study is robust speech feature prediction using mel-LPC to improve the performance of speech recognition in adverse conditions and compares the performance with those standard LPC and MFCC through English dictation system with 14,000 isolated words and 9,000 connected words. The mel-LPC feature prediction is estimated by an optimal value of frequency warping factor that can be estimated from the auto-correlation coefficients and it is computed as the inverse Fourier transform of the power spectrum to generate feature extract vector. Results of the feature extraction are a sequence of 18 mel-LPC coefficients which characteristic of the time-varying spectral properties of the speech signal and these are continuous that can map to discrete vectors in vector quantization codebook. This system is trained by 10 male and 10 female speakers and tested with 200 speakers in noisy and clean environments. Experiments results for various tasks show that with new mel-LPC feature vector system attains isolated and connected word accuracy of 97.5 and 93.2% for male speakers and 96.6 and 92.3% for female speakers with large vocabularies. The result shows that recognition accuracy is relatively higher than LPC and MFCC, respectively.