In this work a novel speech region detection method is proposed that is robust for implementation of speaker verification (SV) systems in practical scenario. Vowel-like region (VLR)s have been found to work better in case of degraded conditions that helps for SV. The proposed method for speech region detection first identifies the VLRs of a given speech utterance and linear prediction (LP) spectrum is computed for those regions. It is observed that the spectral slopes for speech and non-speech regions are quite discriminative which can be useful for classification of the two categories. A study is made for the discrimination of the two classes and a threshold is obtained to categorize into speech or non-speech regions. The proposed approach is evaluated over a text-dependent SV database collected from a practical system developed over online telephone based network. It is observed that the above mentioned approach for robust speech region detection works better than the existing approaches for SV systems in a practical scenario. © 2017 IEEE.